Jump to contentJump to page navigation: previous page [access key p]/next page [access key n]
documentation.suse.com / Optimizing Linux for AMD EPYC™ 9004 Series Processors with SUSE Linux Enterprise 15 SP4
SUSE Linux Enterprise Server 15 SP4

Optimizing Linux for AMD EPYC™ 9004 Series Processors with SUSE Linux Enterprise 15 SP4

SUSE Best Practices
Mel Gorman, Senior Kernel Engineer (SUSE)
Martin Jambor, Tool Chain Developer (SUSE)
Brent Hollingsworth, Engineering Manager (AMD)
SUSE Linux Enterprise 15 SP4
AMD EPYC™ 9004 Series Processors
Date: June 20, 2023
The document at hand provides an overview of both the AMD EPYC™ 9004 Series Classic and AMD EPYC™ 9004 Series Dense Processors. It details how some computational-intensive workloads can be tuned on SUSE Linux Enterprise Server 15 SP4.Disclaimer: Documents published as part of the SUSE Best Practices series have been contributed voluntarily by SUSE employees and third parties. They are meant to serve as examples of how particular actions can be performed. They have been compiled with utmost attention to detail. However, this does not guarantee complete accuracy. SUSE cannot verify that actions described in these documents do what is claimed or whether actions described have unintended consequences. SUSE LLC, its affiliates, the authors, and the translators may not be held liable for possible errors or the consequences thereof.

1 Overview

The AMD EPYC 9004 Series Processor is the latest generation of the AMD64 System-on-Chip (SoC) processor family. It is based on the Zen 4 microarchitecture introduced in 2022, supporting up to 96 cores (192 threads) and 12 memory channels per socket. At the time of writing, 1-socket and 2-socket models are expected to be available from Original Equipment Manufacturers (OEMs) in 2023. In 2023, a new AMD EPYC 9004 Series Dense Processor was launched which is based on a similar architecture to the AMD EPYC 9004 Series Classic Processor supporting up to 128 cores (256 threads). This document provides an overview of the AMD EPYC 9004 Series Classic Processor and how computational-intensive workloads can be tuned on SUSE Linux Enterprise Server 15 SP4. Additional details about the AMD EPYC 9004 Series Dense Processor are provided where appropriate.

2 AMD EPYC 9004 Series Classic Processor architecture

Symmetric multiprocessing (SMP) systems are those that contain two or more physical processing cores. Each core may have two threads if Symmetric multithreading (SMT) is enabled, with some resources being shared between SMT siblings. To minimize access latencies, multiple layers of caches are used with each level being larger but with higher access costs. Cores may share different levels of cache which should be considered when tuning for a workload.

Historically, a single socket contained several cores sharing a hierarchy of caches and memory channels and multiple sockets were connected via a memory interconnect. Modern configurations may have multiple dies as a Multi-Chip Module (MCM) with one set of interconnects within the socket and a separate interconnect for each socket. This means that some CPUs and memory are faster to access than others depending on the distance. This should be considered when tuning for Non-Uniform Memory Architecture (NUMA) as all memory accesses may not reference local memory incurring a variable access penalty.

The 4th Generation AMD EPYC Processor has an MCM design with up to thirteen dies on each package. From a topology point of view, this is significantly different to the 1st Generation AMD EPYC Processor design. However, it is similar to the 3rd Generation AMD EPYC Processor other than the increase in die count. One die is a central IO die through which all off-chip communication passes through. The basic building block of a compute die is an eight-core Core CompleX (CCX) with its own L1-L3 cache hierarchy. Similar to the 3rd Generation AMD EPYC Processor, one Core Complex Die (CCD) consists of one CCX connected via an Infinity Link to the IO die, as opposed to two CCXs used in the 2nd Generation AMD EPYC Processor. This allows direct communication within a CCD instead of using the IO link maintaining reduced communication and memory access latency. A 96-core 4th Generation AMD EPYC Processor socket therefore consists of 12 CCDs consisting of 12 CCXs (containing 8 cores each) or 96 cores in total (192 threads with SMP enabled) with one additional IO die for 13 dies in total. This is a large increase in both the core count and number of memory channels relative to the 3rd Generation AMD EPYC Processor.

Both the 3rd and 4th Generation AMD EPYC Processors potentially have a larger L3 cache. In a standard configuration, a 4th Generation AMD EPYC Processor has 32MB L3 cache. Some CPU chips may also include an AMD V-Cache expansion that can triple the size of the L3 cache. This potentially provides a major performance boost to applications as more active data can be stored in low-latency cache. The exact performance impact is variable, but any memory-intensive workload should benefit from having a lower average memory access latency because of a larger cache.

Communication between the chip and memory happens via the IO die. Each CCD has one dedicated Infinity Fabric link to the die and one memory channel per CCD located on the die. The practical consequence of this architecture versus the 1st Generation AMD EPYC Processor is that the topology is simpler. The first generation had separate memory channels per die and links between dies giving two levels of NUMA distance within a single socket and a third distance when communicating between sockets. This meant that a two-socket machine for EPYC had 4 NUMA nodes (3 levels of NUMA distance). The 2nd Generation AMD EPYC Processor has only 2 NUMA nodes (2 levels of NUMA distance) which makes it easier to tune and optimize. The NUMA distances are the same for the 3rd and 4th Generation AMD EPYC Processors.

The IO die has a total of 12 memory controllers supporting DDR5 Dual Inline Memory Modules (DIMMs) with the maximum supported speed expected to be DDR-5200 at the time of writing. This implies a peak channel bandwidth of 40.6 GB/sec or 487.2 GB/sec total throughput across a socket. The exact bandwidth depends on the DIMMs selected, the number of memory channels populated, how cache is used and the efficiency of the application. Where possible, all memory channels should have a DIMM installed to maximize memory bandwidth.

While the topologies and basic layout is similar between the 3rd and 4th Generation AMD EPYC Processors, there are several micro-architectural differences. The Instructions Per Cycle (IPC) has improved by 13% on average across a selected range of workloads, although the exact improvement is workload-dependent. The improvements are due to a variety of factors including a larger L2 cache, improvements in branch prediction, the execution engine, the front-end fetching/decoding of instructions and additional instructions such as supporting AVX-512. The degree to which these changes affect performance varies between applications.

Power management on the links is careful to minimize the amount of power required. If the links are idle, the power may be used to boost the frequency of individual cores. Hence, minimizing access is not only important from a memory access latency point of view, but it also has an impact on the speed of individual cores.

There are 128 IO lanes supporting PCIe Gen 5.0 per socket. Lanes can be used as Infinity links, PCI Express links or SATA links with a limit of 32 SATA links. The exact number of PCIe 4.0, PCIe 5.0 and configuration links vary by chip and motherboard. This allows very large IO configurations and a high degree of flexibility, given that either IO bandwidth or the bandwidth between sockets can be optimized, depending on the OEM requirements. The most likely configuration is that the number of PCIe links will be the same for 1- and 2-socket machines, given that some lanes per socket will be used for inter-socket communication. While some links must be used for inter-socket communication, adding a socket does not compromise the number of available IO channels. The exact configuration used depends on the platform.

3 AMD EPYC 9004 Series Classic Processor topology

Figure 1, “AMD EPYC 9004 Series Classic Processor Topology” below shows the topology of an example machine with a fully populated memory configuration generated by the lstopo tool.

AMD EPYC 9004 Series Classic Processor Topology
Figure 1: AMD EPYC 9004 Series Classic Processor Topology

This tool is part of the hwloc package. The two packages correspond to each socket. The CCXs consisting of 8 cores (16 threads) each should be clear, as each CCX has one L3 cache and each socket has 12 CCXs resulting in 96 cores (192 threads). Not obvious are the links to the IO die, but the IO die should be taken into account when splitting a workload to optimize bandwidth to memory. In this example, the IO channels are not heavily used, but the focus will be on CPU and memory-intensive loads. If optimizing for IO, it is recommended that, where possible, the workload is located on the nodes local to the IO channel.

The computer output below shows a conventional view of the topology using the numactl tool which is slightly edited for clarity. The CPU IDs that map to each node are reported on the node X cpus: lines. They note the NUMA distances on the table at the bottom of the computer output. Node 0 and node 1 are a distance of 32 apart as they are on separate sockets. The distance is not a guarantee of the access latency, it is an estimate of the relative difference. The general interpretation of this distance would suggest that a remote node is 3.2 times longer than a local memory access but the actual latency cost can be different.

epyc:~ # numactl --hardware
node 0 cpus: 0 .. 95 192 .. 287
node 0 size: 386740 MB
node 0 free: 369215 MB
node 1 cpus: 96 .. 191 288 .. 383
node 1 size: 386771 MB
node 1 free: 371177 MB
node distances:
node   0   1
  0:  10  32
  1:  32  10

Note that the two sockets displayed are masking some details. There are multiple CCDs and multiple channels meaning that there are slight differences in access latency even to local memory. If an application is so sensitive to latency that it needs to be aware of the precise relative distances, then the Nodes Per Socket (NPS) value can be adjusted in the BIOS. If adjusted, numactl will show additional nodes and the relative distances between them.

Finally, the cache topology can be discovered in a variety of fashions. In addition to lstopo which can provide the information, the level, size and ID of CPUs that share cache can be identified from the files under /sys/devices/system/cpu/cpuN/cache.

4 AMD EPYC 9004 Series Dense Processor

The AMD EPYC 9004 Series Dense Processor launched in 2023. While the fundamental microarchitecture is based on the Zen 4 compute core, there are some important differences between it and the AMD EPYC 9004 Series Classic Processors. Both processors are socket-compatible, have the same number of memory channels and the same number of I/O lanes. This means that the processors may be used interchangeably on the same platform with the same limitation that dual-socket configurations must use identical processors. Both processors use the same Instruction Set Architecture (ISA). This means that code optimized for one processor will run without modification on the other.

Despite the compatible ISA, the processors are physically different using a manufacturing process focused on increased density for both the CPU core and the physical cache. The L1 and L2 caches have the same capacity. The L3 cache, however, is half the capacity of the AMD EPYC 9004 Series Classic Processor with the space used for additional CCDs. The basic CCX structure for both the AMD EPYC 9004 Series Dense and 9004 Series Classic processor is similar but each CCD for the AMD EPYC 9004 Series Dense has 2 CCXs per CCD instead of 1. While the AMD EPYC 9004 Series Classic has 12 CCDs with 1 CCX each within a socket, the AMD EPYC 9004 Series Dense processor has 8 CCDs, each with 2 CCXs. This increases the maximum number of its cores per socket from 96 cores to 128. Finally, the Thermal Design Points (TDPs) differ for the AMD EPYC 9004 Series Dense processor, with different frequency scaling limits and generally a lower peak frequency. While each individual core may achieve less peak performance than the AMD EPYC 9004 Series Classic Processor, the total peak compute throughput available is higher due to the increased number of cores.

The intended use case and workloads determine which processor is superior. The key advantage of the AMD EPYC 9004 Series Dense Processor is packing more cores within the same socket. This may benefit Cloud or HyperScale environments in that more containers or virtual machines can use uncontested CPUs for their workloads within the same physical machine. As a result, physical space in data centers can potentially be reduced. It may also benefit some HPC workloads that are primarily CPU and memory bound. For example, some HPC workloads scale to the number of available cores working on data sets that are too large to fit into a typical cache. For such workloads, the AMD EPYC 9004 Series Dense Processor may be ideal.

5 Memory and CPU binding

NUMA is a scalable memory architecture for multiprocessor systems that can reduce contention on a memory channel. A full discussion on tuning for NUMA is beyond the scope for this document. But the document A NUMA API for Linux at http://developer.amd.com/wordpress/media/2012/10/LibNUMA-WP-fv1.pdf provides a valuable introduction.

The default policy for programs is the local policy. A program which calls malloc() or mmap() reserves virtual address space but does not immediately allocate physical memory. The physical memory is allocated the first time the address is accessed by any thread and, if possible, the memory will be local to the accessing CPU. If the mapping is of a file, the first access may have occurred at any time in the past so there are no guarantees about locality.

When memory is allocated to a node, it is less likely to move if a thread changes to a CPU on another node or if multiple programs are remote-accessing the data unless Automatic NUMA Balancing (NUMAB) is enabled. When NUMAB is enabled, unbound process accesses are sampled. If there are enough remote accesses, then the data will be migrated to local memory. This mechanism is not perfect and incurs overhead of its own. This can be important for performance for thread and process migrations between nodes to be minimized and for memory placement to be carefully considered and tuned.

The taskset tool is used to set or get the CPU affinity for new or existing processes. An example usage is to confine a new process to CPUs local to one node. Where possible, local memory will be used. But if the total required memory is larger than the node, then remote memory can still be used. In such configurations, it is recommended to size the workload such that it fits in the node. This avoids that any of the data is being paged out when kswapd wakes to reclaim memory from the local node.

numactl controls both memory and CPU policies for processes that it launches and can modify existing processes. In many respects, the parameters are easier to specify than taskset. For example, it can bind a task to all CPUs on a specified node instead of having to specify individual CPUs with taskset. Most importantly, it can set the memory allocation policy without requiring application awareness.

Using policies, a preferred node can be specified where the task will use that node if memory is available. This is typically used in combination with binding the task to CPUs on that node. If a workload's memory requirements are larger than a single node and predictable performance is required, then the interleave policy will round-robin allocations from allowed nodes. This gives suboptimal but predictable access latencies to main memory. More importantly, interleaving reduces the probability that the operating system (OS) will need to reclaim any data belonging to a large task.

Further improvements can be made to access latencies by binding a workload to a single CCD within a node. As L3 caches are shared within a CCD on both the 3rd and 4th Generation AMD EPYC Processors, binding a workload to a CCD avoids L3 cache misses caused by workload migration. This is an important difference from the 2nd Generation AMD EPYC Processor which favored binding within a CCX.

In most respects, the guidance for optimal bindings for cache and nodes remains the same between the 3rd and 4th Generation AMD EPYC Processors. However, with SUSE Linux Enterprise 15 SP4, the necessity to bind specifically to the L3 cache for optimal performance is relaxed. The CPU scheduler in SUSE Linux Enterprise 15 SP4 has superior knowledge of the cache topology of all generations of the AMD EPYC Processors and how to balance load between CPU caches, NUMA nodes and memory channels.

Note: CPU Scheduler Awareness of Cache Topology

With SUSE Linux Enterprise 15 SP4 having superior knowledge of the CPU cache topology and how to balance load, tuning specifically has a smaller impact to performance for a given workload. This is not a limitation of the operating system. It is a side-effect of the baseline performance being improved on AMD EPYC Processors in general.

See examples below on how taskset and numactl can be used to start commands bound to different CPUs depending on the topology.

# Run a command bound to CPU 1
epyc:~ # taskset -c 1 [command]
# Run a command bound to CPUs belonging to node 0
epyc:~ # taskset -c `cat /sys/devices/system/node/node0/cpulist` [command]
# Run a command bound to CPUs belonging to nodes 0 and 1
epyc:~ # numactl –cpunodebind=0,1 [command]
# Run a command bound to CPUs that share L3 cache with cpu 1
epyc:~ # taskset -c `cat /sys/devices/system/cpu/cpu1/cache/index3/shared_cpu_list` [command]

5.1 Tuning for local access without binding

The ability to use local memory where possible and remote memory if necessary is valuable. But there are cases where it is imperative that local memory always be used. If this is the case, the first priority is to bind the task to that node. If that is not possible, then the command sysctl vm.zone_reclaim_mode=1 can be used to aggressively reclaim memory if local memory is not available.

Note: Potential Hazard with vm.zone_reclaim_mode

While this option is good from a locality perspective, it can incur high costs because of stalls related to reclaim and the possibility that data from the task will be reclaimed. Treat this option with a high degree of caution and testing.

5.2 Hazards with CPU binding

There are three major hazards to consider with CPU binding.

The first to watch for is remote memory nodes being used when the process is not allowed to run on CPUs local to that node. The scenarios when this can occur are outside the scope of this paper. However, a common reason is an IO-bound thread communicating with a kernel IO thread on a remote node bound to the IO controller. In such a setup, the data buffers managed by the application are stored in remote memory incurring an access cost for the IO.

While tasks may be bound to CPUs, the resources they are accessing, such as network or storage devices, may not have interrupts routed locally. irqbalance generally makes good decisions. But in cases where the network or IO is extremely high-performance or the application has very low latency requirements, it may be necessary to disable irqbalance using systemctl. When that is done, the IRQs for the target device need to be routed manually to CPUs local to the target workload for optimal performance.

The second is that guides about CPU binding tend to focus on binding to a single CPU. This is not always optimal when the task communicates with other threads, as fixed bindings potentially miss an opportunity for the processes to use idle CPUs sharing a common cache. This is particularly true when dispatching IO, be it to disk or a network interface, where a task may benefit from being able to migrate close to the related threads. It also applies to pipeline-based communicating threads for a computational workload. Hence, focus initially on binding to CPUs sharing L3 cache. Then consider whether to bind based on an L1/L2 cache or a single CPU using the primary metric of the workload to establish whether the tuning is appropriate.

The final hazard is similar: if many tasks are bound to a smaller set of CPUs, then the subset of CPUs could be oversaturated even though there is spare CPU capacity available.

5.3 CPUsets and memory control groups

CPUsets are ideal when multiple workloads must be isolated on a machine in a predictable fashion. CPUsets allow a machine to be partitioned into subsets. These sets may overlap, and in that case they suffer from similar problems as CPU affinities. In the event there is no overlap, they can be switched to exclusive mode which treats them completely in isolation with relatively little overhead. Similarly, they are well suited when a primary workload must be protected from interference because of low-priority tasks. In such cases the low-priority tasks can be placed in a CPUset. The caveat with CPUsets is that the overhead is higher than using scheduler and memory policies. Ordinarily, the accounting code for CPUsets is completely disabled. But when a single CPUset is created, there is a second layer of checks against scheduler and memory policies.

Similarly, memcg can be used to limit the amount of memory that can be used by a set of processes. When the limits are exceeded, the memory will be reclaimed by tasks within memcg directly without interfering with any other tasks. This is ideal for ensuring there is no inference between two or more sets of tasks. Similar to CPUsets, there is some management overhead incurred. This means, if tasks can simply be isolated on a NUMA boundary, then this is preferred from a performance perspective. The major hazard is that, if the limits are exceeded, then the processes directly stall to reclaim the memory which can incur significant latencies.


Without memcg, when memory gets low, the global reclaim daemon does work in the background and if it reclaims quickly enough, no stalls are incurred. When using memcg, observe the allocstall counter in /proc/vmstat as this can detect early if stalling is a problem.

6 High performance storage devices and interrupt affinity

High performance storage devices like Non-Volatile Memory Express (NVMe) or Serial Attached SCSI (SAS) controller are designed to take advantage of parallel IO submission. These devices typically support a large number of submit and receive queues, which are tied to MSI-X interrupts. Ideally, these devices should provide as many MSI-X vectors as there are CPUs in the system. To achieve the best performance, each MSI-X vector should be assigned to an individual CPU.

7 Automatic NUMA balancing

Automatic NUMA Balancing (NUMAB) is a feature that identifies and relocates pages that are being accessed remotely for applications that are not NUMA-aware. There are cases where it is impractical or impossible to specify policies. In such cases, the balancing should be sufficient for throughput-sensitive workloads, but on occasion, NUMAB may be considered hazardous as it incurs a cost. Under ideal conditions, an application is NUMA aware and uses memory policies to control what memory is accessed and NUMAB simply ignores such regions. However, even if an application does not use memory policies, it is possible that the application still accesses mostly local memory and NUMA adds overhead confirming that accesses are local which is an unnecessary cost. For latency-sensitive workloads, the sampling for NUMA balancing may be too unpredictable and would prefer to incur the remote access cost or interleave memory instead of using NUMA. The final corner case where NUMA balancing is a hazard happens when the number of runnable tasks always exceeds the number of CPUs in a single node. In this case, the load balancer (and potentially affine wakes) may pull tasks away from the preferred node as identified by Automatic NUMA balancing resulting in excessive sampling and migrations.

If the workloads can be manually optimized with policies, then consider disabling automatic NUMA balancing by specifying numa_balancing=disable on the kernel command line or via sysctl kernel.numa_balancing. The same applies if it is known that the application is mostly accessing local memory.

Note: Changes to Automatic NUMA Balancing

While a disconnect between CPU Scheduler and NUMA Balancing placement decisions still potentially exists in SUSE Linux Enterprise 15 SP4 when the machine is heavily overloaded, the impact is much reduced relative to previous releases for most scenarios. The placement decisions made by the CPU Scheduler and NUMA Balancing are now coupled. Situations where the CPU scheduler and NUMA Balancing make opposing decisions are relatively rare.

8 Evaluating workloads

The first and foremost step when evaluating how a workload should be tuned is to establish a primary metric such as latency, throughput, operations per second or elapsed time. When each tuning step is considered or applied, it is critical that the primary metric be examined before conducting any further analysis. This is to avoid intensive focus on a relatively wrong bottleneck. Make sure that the metric is measured multiple times to ensure that the result is reproducible and reliable within reasonable boundaries. When that is established, analyze how the workload is using different system resources to determine what area should be the focus. The focus in this paper is on how CPU and memory is used. But other evaluations may need to consider the IO subsystem, network subsystem, system call interfaces, external libraries, etc. The methodologies that can be employed to conduct this are outside the scope of this paper. But the book Systems Performance: Enterprise and the Cloud by Brendan Gregg (see http://www.brendangregg.com/systems-performance-2nd-edition-book.html) is a recommended primer on the subject.

8.1 CPU utilization and saturation

Decisions on whether to bind a workload to a subset of CPUs require that the CPU utilization and any saturation risk is known. Both the ps and pidstat commands can be used to sample the number of threads in a system. Typically, pidstat yields more useful information with the important exception of the run state. A system may have many threads, but if they are idle, they are not contributing to utilization. The mpstat command can report the utilization of each CPU in the system.

High utilization of a small subset of CPUs may be indicative of a single-threaded workload that is pushing the CPU to the limits and may indicate a bottleneck. Conversely, low utilization may indicate a task that is not CPU-bound, is idling frequently or is migrating excessively. While each workload is different, load utilization of CPUs may show a workload that can run on a subset of CPUs to reduce latencies because of either migrations or remote accesses. When utilization is high, it is important to determine if the system could be saturated. The vmstat tool reports the number of runnable tasks waiting for a CPU in the r column where any value over 1 indicates that wakeup latencies may be incurred. While the exact wakeup latency can be calculated using trace points, knowing that there are tasks queued is an important step. If a system is saturated, it may be possible to tune the workload to use fewer threads.

Overall, the initial intent should be to use CPUs from as few NUMA nodes as possible to reduce access latency. However, there are exceptions. The AMD EPYC 9004 Series Processor has a large number of high-speed memory channels to main memory, so consider the workload thread activity. If they are cooperating threads or sharing data, isolate them on as few nodes as possible to minimize cross-node memory accesses. If the threads are completely independent with no shared data, it may be best to isolate them on a subset of CPUs from each node. This is to maximize the number of available memory channels and throughput to main memory. For some computational workloads, it may be possible to use hybrid models such as MPI for parallelization across nodes and OpenMP for threads within nodes.

Note: Updating tuning for AMD EPYC 9004 Series Processor

It is expected that tuning based on the AMD EPYC 7003 Series Processor will also usually perform optimally on AMD EPYC 9004 series. The main consideration is to account for potential differences in L3 cache sizes because of AMD V-Cache if workloads are tuned specifically for cache size. Also, keep in mind that CPU bindings based on caches may potentially be relaxed on SUSE Linux Enterprise 15 SP4.

8.2 Transparent Huge Pages

Huge pages are a feature that can improve performance in many cases. This is achieved by reducing the number of page faults, the cost of translating virtual addresses to physical addresses because of fewer layers in the page table and being able to cache translations for a larger portion of memory. Transparent Huge Pages (THP) is supported for private anonymous memory that automatically backs mappings with huge pages where anonymous memory could be allocated as heap, malloc(), mmap(MAP_ANONYMOUS), etc. There is also support for using THP pages backed by tmpfs which can be configured at mount time using the huge= mount option. While the THP feature has existed for a long time, it has evolved significantly.

Many tuning guides recommend disabling THP because of problems with early implementations. Specifically, when the machine was running for long enough, the use of THP could incur severe latencies and could aggressively reclaim memory in certain circumstances. These problems were resolved by the time SUSE Linux Enterprise Server 15 SP2 was released, and this is still the case for SUSE Linux Enterprise Server 15 SP4. This means there are no good grounds for automatically disabling THP because of severe latency issues without measuring the impact. However, there are exceptions that are worth considering for specific workloads.

Some high-end in-memory databases and other applications aggressively use mprotect() to ensure that unprivileged data is never leaked. If these protections are at the base page granularity, then there may be many THP splits and rebuilds that incur overhead. It can be identified if this is a potential problem by using strace or perf trace to detect the frequency and granularity of the system call. If they are high-frequency, consider disabling THP. It can also be sometimes inferred from observing the thp_split and thp_collapse_alloc counters in /proc/vmstat.

Workloads that sparsely address large mappings may have a higher memory footprint when using THP. This could result in premature reclaim or fallback to remote nodes. An example would be HPC workloads operating on large sparse matrices. If memory usage is much higher than expected, compare memory usage with and without THP to decide if the trade-off is not worthwhile. This may be critical on AMD EPYC 7003 and 9004 Series Processor given that any spillover will congest the Infinity links and potentially cause cores to run at a lower frequency.

Note: Sparsely addressed memory

This is specific to sparsely addressed memory. A secondary hint for this case may be that the application primarily uses large mappings with a much higher Virtual Size (VSZ, see Section 8.4, “Memory utilization and saturation”) than Resident Set Size (RSS). Applications which densely address memory benefit from the use of THP by achieving greater bandwidth to memory.

Parallelized workloads that operate on shared buffers with thread counts exceeding the number of available CPUs on a single node may experience a slowdown with THP if the granularity of partitioning is not aligned to the huge page. The problem is that if a large shared buffer is partitioned on a 4K boundary, then false sharing may occur whereby one thread accesses a huge page locally and other threads access it remotely. If this situation is encountered, the granularity of sharing should be increased to the THP size. But if that is not possible, disabling THP is an option.

Applications that are extremely latency-sensitive or must always perform in a deterministic fashion can be hindered by THP. While there are fewer faults, the time for each fault is higher as memory must be allocated and cleared before being visible. The increase in fault times may be in the microsecond granularity. Ensure this is a relevant problem as it typically only applies to extremely latency-sensitive applications. The secondary problem is that a kernel daemon periodically scans a process looking for contiguous regions that can be backed by huge pages. When creating a huge page, there is a window during which that memory cannot be accessed by the application and new mappings cannot be created until the operation is complete. This can be identified as a problem with thread-intensive applications that frequently allocate memory. In this case, consider effectively disabling khugepaged by setting a large value in /sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs. This will still allow THP to be used opportunistically while avoiding stalls when calling malloc() or mmap().

THP can be disabled. To do so, specify transparent_hugepage=disable on the kernel command line, at runtime via /sys/kernel/mm/transparent_hugepage/enabled or on a per-process basis by using a wrapper to execute the workload that calls prctl(PR_SET_THP_DISABLE).

8.3 User/kernel footprint

Assuming an application is mostly CPU- or memory-bound, it is useful to determine if the footprint is primarily in user space or kernel space. This gives a hint where tuning should be focused. The percentage of CPU time can be measured on a coarse-grained fashion using vmstat or a fine-grained fashion using mpstat. If an application is mostly spending time in user space, then the focus should be on tuning the application itself. If the application is spending time in the kernel, then it should be determined which subsystem dominates. The strace or perf trace commands can measure the type, frequency and duration of system calls as they are the primary reasons an application spends time within the kernel. In some cases, an application may be tuned or modified to reduce the frequency and duration of system calls. In other cases, a profile is required to identify which portions of the kernel are most relevant as a target for tuning.

8.4 Memory utilization and saturation

The traditional means of measuring memory utilization of a workload is to examine the Virtual Size (VSZ) and Resident Set Size (RSS). This can be done by using either the ps or pidstat tool. This is a reasonable first step but is potentially misleading when shared memory is used and multiple processes are examined. VSZ is simply a measure of memory space reservation and is not necessarily used. RSS may be double accounted if it is a shared segment between multiple processes. The file /proc/pid/maps can be used to identify all segments used and whether they are private or shared. The file /proc/pid/smaps and /proc/pid/smaps_rollup reveals more detailed information including the Proportional Set Size (PSS). PSS is an estimate of RSS except it is divided between the number of processes mapping that segment, which can give a more accurate estimate of utilization. Note that the smaps and smaps_rollup files are very expensive to read and should not be monitored at a high frequency. This is especially the case if workloads are using large amounts of address space, many threads or both. Finally, the Working Set Size (WSS) is the amount of active memory required to complete computations during an arbitrary phase of a program's execution. It is not a value that can be trivially measured. But conceptually it is useful as the interaction between WSS relative to available memory affects memory residency and page fault rates.

On NUMA systems, the first saturation point is a node overflow when the local policy is in effect. Given no binding of memory, when a node is filled, a remote node’s memory will be used transparently and background reclaim will take place on the local node. Two consequences of this are that remote access penalties will be used and old memory from the local node will be reclaimed. If the WSS of the application exceeds the size of a local node, then paging and re-faults may be incurred.

The first item to identify is whether a remote node overflow occurred, which is accounted for in /proc/vmstat as the numa_hit, numa_miss, numa_foreign, numa_interleave, numa_local and numa_other counters:

  • numa_hit is incremented when an allocation uses the preferred node where preferred may be either a local node or one specified by a memory policy.

  • numa_miss is incremented when an alternative node is used to satisfy an allocation.

  • numa_foreign is rarely useful but is accounted against a node that was preferred. It is a subtle distinction from numa_miss that is rarely useful.

  • numa_interleave is incremented when an interleave policy was used to select allowed nodes in a round-robin fashion.

  • numa_local increments when a local node is used for an allocation regardless of policy.

  • numa_other is used when a remote node is used for an allocation regardless of policy.

For the local memory policy, the numa_hit and numa_miss counters are the most important to pay attention to. An application that is allocating memory that starts incrementing the numa_miss implies that the first level of saturation has been reached. If monitoring the proc is undesirable, then the numastat provides the same information. If this is observed on the AMD EPYC 9004 Series Processor, it may be valuable to bind the application to nodes that represent dies on a single socket. If the ratio of hits to misses is close to 1, consider an evaluation of the interleave policy to avoid unnecessary reclaim.

Note: NUMA statistics

These NUMA statistics only apply at the time a physical page is allocated and are not related to the reference behavior of the workload. For example, if a task running on node 0 allocates memory local to node 0, then it will be accounted for as a node_hit in the statistics. However, if the memory is shared with a task running on node 1, all the accesses may be remote, which is a miss from the perspective of the hardware but not accounted for in /proc/vmstat. Detecting remote and local accesses at the hardware level requires using the hardware's Performance Management Unit to detect.

When the first saturation point is reached, reclaim will be active. This can be observed by monitoring the pgscan_kswapd and pgsteal_kswapd counters in /proc/vmstat. If this is matched with an increase in major faults or minor faults, then it may be indicative of severe thrashing. In this case, the interleave policy should be considered. An ideal tuning option is to identify if shared memory is the source of the usage. If this is the case, then interleave the shared memory segments. This can be done in some circumstances using numactl or by modifying the application directly.

More severe saturation is observed if the pgscan_direct and pgsteal_direct counters are also increasing. These counters indicate that the application is stalling while memory is being reclaimed. If the application was bound to individual nodes, increasing the number of available nodes will alleviate the pressure. If the application is unbound, it indicates that the WSS of the workload exceeds all available memory. It can only be alleviated by tuning the application to use less memory or increasing the amount of RAM available.

A more generalized view of resource pressure for CPU, memory and IO can be measured using the kernel Pressure Stall Information feature enabled with the command line psi=1. When enabled, proc files under /proc/pressure show if some or all active tasks were stalled recently contending on a resource. This information is not always available in production. But if the information is available, the memory pressure information may be used to guide whether a deeper analysis is necessary and which resource is the bottleneck.

As before, whether to use memory nodes from one socket or two sockets depends on the application. If the individual processes are independent, either socket can be used. Where possible, keep communicating processes on the same socket to maximize memory throughput while minimizing the socket interconnect traffic.

8.5 Other resources

The analysis of other resources is outside the scope of this paper. However, a common scenario is that an application is IO-bound. A superficial check can be made using the vmstat tool. This tool checks what percentage of CPU time is spent idle combined with the number of processes that are blocked and the values in the bi and bo columns. Similarly, if PSI is enabled, then the IO pressure file will show whether some or all active tasks are losing time because of lack of resources. Further analysis is required to determine if an application is IO rather than CPU- or memory-bound. However, this is a sufficient check to start with.

9 Power management

Modern CPUs balance power consumption and performance through Performance States (P-States). Low utilization workloads may use lower P-States to conserve power while still achieving acceptable performance. When a CPU is idle, lower power idle states (C-States) can be selected to further conserve power. However, this comes with higher exit latencies when lower power states are selected. It is further complicated by the fact that, if individual cores are idle and running at low power, the additional power can be used to boost the performance of active cores. This means this scenario is not a straightforward balance between power consumption and performance. More complexity is added on the AMD EPYC 7003 and 9004 Series Processors whereby spare power may be used to boost either cores or the Infinity links.

The 4th Generation AMD EPYC Processor provides SenseMI. This technology, among other capabilities, enables CPUs to make adjustments to voltage and frequency depending on the historical state of the CPU. There is a latency penalty when switching P-States, but the AMD EPYC 9004 Series Processor is capable of making fine-grained adjustments to reduce the likelihood that the latency is a bottleneck. On SUSE Linux Enterprise Server, the AMD EPYC 9004 Series Processor uses the acpi_cpufreq driver. This allows P-states to be configured to match requested performance. However, this is limited in terms of the full capabilities of the hardware. It cannot boost the frequency beyond the maximum stated frequencies, and if a target is specified, then the highest frequency below the target will be used. A special case is if the governor is set to performance. In this situation the hardware will use the highest available frequency in an attempt to work quickly and then return to idle.

What should be determined is whether power management is likely to be a factor for a workload. A single thread workload that is CPU-bound is likely to run at the highest frequency on a single core. Lastly, a workload that does not communicate heavily with other processes and is mostly CPU-bound is unlikely to experience any side effects because of power management. The exceptions are when load balancing moves tasks away from active CPUs if there is a compute imbalance between NUMA nodes or the machine is heavily overloaded.

The workloads that are most likely to be affected by power management are those that:

  • synchronously communicate between multiple threads.

  • idle frequently.

  • have low CPU utilization overall.

It will be further compounded if the threads are sensitive to wakeup latency.

Power management is critical, not only for power savings, but because power saved from idling inactive cores can be used to boost the performance of active cores. On the other side, low utilization tasks may take longer to complete if the task is not active long enough for the CPU to run at a high frequency. In some cases, problems can be avoided by configuring a workload to use the minimum number of CPUs necessary for its active tasks. Deciding that means monitoring the power state of CPUs.

The P-State and C-State of each CPU can be examined using the turbostat utility. The computer output below shows an example, slightly edited to fit the page, where a workload is busy on CPU 0 and other workloads are idle. A useful exercise is to start a workload and monitor the output of turbostat paying close attention to CPUs that have moderate utilization and running at a lower frequency. If the workload is latency-sensitive, it is grounds for either minimizing the number of CPUs available to the workload or configuring power management.

Pac. Die  Core CPU Avg_M Busy% Bzy_M TSC_M IRQ   POLL   C1 C2  POLL% C1%  C2%
-    -    -    -   1516  44.16 3432  1198  1.65  340109 90 706 6937  0.00 0.00
0    0    0    0   463   13.24 3498  1200  0.26  2936   0  0   1406  0.00 0.00
0    0    0    192 1     0.04  3499  1200  0.10  27     0  0   21    0.00 0.00
0    0    1    1   3189  91.16 3498  1200  1.52  1792   0  0   18    0.00 0.00
0    0    1    193 317   9.07  3498  1200  0.55  203    0  0   18    0.00 0.00
0    0    2    2   3436  98.22 3498  1200  1.40  1905   0  0   3     0.00 0.00
0    0    2    194 78    2.22  3500  1200  0.92  84     0  0   29    0.00 0.00

If tuning CPU frequency management is appropriate, the following actions can be taken to set the management policy to performance using the cpupower utility:

epyc:~# cpupower frequency-set -g performance
Setting cpu: 0
Setting cpu: 1
Setting cpu: 2

Persisting it across reboots can be done via a local init script, via udev or via one-shot systemd service file if necessary. Note that turbostat will still show that idling CPUs use a low frequency. The impact of the policy is that the highest P-State will be used as soon as possible when the CPU is active. In some cases, a latency bottleneck will occur because of a CPU exiting idle. If this is identified on the AMD EPYC 9004 Series Processor, restrict the C-state by specifying processor.max_cstate=2 if lower P-States exist on the kernel command line. This will prevent CPUs from entering lower C-states. The availability of P-states can be determined with cpupower idle-info. It is expected on the AMD EPYC 9004 Series Processor that the exit latency from C1 is very low. But by allowing C2, it reduces interference from the idle loop injecting micro-operations into the pipeline and should be the best state overall. It is also possible to set the max idle state on individual cores using cpupower idle-set. If SMT is enabled, the idle state should be set on both siblings.

10 Security mitigation

On occasion, a security fix is applied to a distribution that has a performance impact. The most notable example is Meltdown and multiple variants of Spectre but includes others such as ForeShadow (L1TF). The AMD EPYC 9004 Series Processor is immune to the Meltdown variant and page table isolation is never active. However, it is vulnerable to a subset of Spectre variants although retbleed is a notable exception. The following table lists all security vulnerabilities that affect the 4th Generation AMD EPYC Processor. In addition, it specifies which mitigations are enabled by default for SUSE Linux Enterprise Server 15 SP4.

Table 1: Security mitigations for AMD EPYC 9004 Series Processors
VulnerabilityAffected Mitigations

ITLB Multihit












MMIO Stale Data






Speculative Store Bypass


prctl and seccomp policy

Spectre v1


usercopy/swapgs barriers and __user pointer sanitization

Spectre v2


Retpoline, RSB filling, and conditional IBPB, IBRS_FW, and STIBP




TSX Async Abort



If it can be guaranteed that the server is in a trusted environment running only known code that is not malicious, the mitigations=off parameter can be specified on the kernel command line. This option disables all security mitigations and may improve performance in some cases. However, at the time of writing and in most cases, the gain on an AMD EPYC 9004 Series Processor is marginal when compared to other CPUs. Evaluate carefully whether the gain is worth the risk and if unsure, leave the mitigations enabled.

11 Hardware-based profiling

The AMD EPYC 9004 Series Processor has extensive Performance Monitoring Unit (PMU) capabilities. Advanced monitoring of a workload can be conducted via the perf. The command supports a range of hardware events including cycles, L1 cache access/misses, TLB access/misses, retired branch instructions and mispredicted branches. To identify what subsystem may be worth tuning in the OS, the most useful invocation is perf record -a -e cycles sleep 30. This captures 30 seconds of data for the entire system. You can also call perf record -e cycles command to gather a profile of a given workload. Specific information on the OS can be gathered through tracepoints or creating probe points with perf or trace-cmd. But the details on how to conduct such analyses are beyond the scope of this paper.

12 Compiler selection

SUSE Linux Enterprise ships with multiple versions of GCC. SUSE Linux Enterprise 15 SP4 service packs ship with GCC 7 which at the time of writing is GCC 7.5.0 with the package version 7-3.9.1. The intention is to avoid unintended consequences when porting code that may affect the building and operation of applications. The GCC 7 development originally started in 2016, with a branch created in 2017 and GCC 7.5 released in 2019. This means that the system compiler has no awareness of the AMD EPYC 7002 or later Series processors.

Fortunately, the add-on Developer Tools Module includes additional compilers with the latest version currently based on GCC 11.2.1. This compiler is capable of generating optimized code targeted at the 3rd Generation AMD EPYC Processor using the znver3 target. It also provides additional support for OpenMP 5.0. Unlike the system compiler, the major version of GCC shipped with the Developer Tools Module can change during the lifetime of the product. It is expected that GCC 12 will be included in future releases for generating optimized code for the 4th Generation AMD EPYC Processor. Unfortunately, at the time of writing, there is not a version of GCC available optimized for the AMD EPYC 9004 Series Processor specifically.

The OS packages are built against a generic target. However, where applications and benchmarks can be rebuilt from source, the minimum option should be -march=znver3 for GCC 11 and later versions of GCC.

Further information on how to install the Developer Tools Module and how to build optimized versions of applications can be found in the guide Advanced optimization and new capabilities of GCC 11.

13 Candidate workloads

The workloads that will benefit most from the 4th Generation AMD EPYC Processor architecture are those that can be parallelized and are either memory or IO-bound. This is particularly true for workloads that are NUMA friendly. They can be parallelized, and each thread can operate independently for most of the workload's lifetime. For memory-bound workloads, the primary benefit will be taking advantage of the high bandwidth available on each channel. For IO-bound workloads, the primary benefit will be realized when there are multiple storage devices, each of which is connected to the node local to a task issuing IO.

13.1 Test setup

The following sections will demonstrate how an OpenMP and MPI workload can be configured and tuned on an AMD EPYC 9004 Series Processor reference platform. The system has two processors, each with 96 cores and SMT enabled for a total of 192 cores (384 logical CPUs). The peak bandwidth available to the machine depends on the type of DIMMs installed and how the DIMM slots are populated. Note that the peak theoretical transfer speed is rarely reached in practice, given that it can be affected by the mix of reads/writes and the location and temporal proximity of memory locations accessed.

Table 2: Test setup


2x AMD EPYC 9654


AMD Engineering Sample Platform


Samsung SSD 983 DCT 960GB


SUSE Linux Enterprise Server 15 SP4

Memory Type

24x 32GB DDR5

Memory Interleaving


Memory Speed

4800 MT/sec

Kernel command line


13.2 Test workload: STREAM

STREAM is a memory bandwidth benchmark created by Dr. John D. McCalpin from the University of Virginia (for more information, see https://www.cs.virginia.edu/stream/). It can be used to measure bandwidth of each cache level and bandwidth to main memory while calculating four basic vector operations. Each operation can exhibit different throughputs to main memory depending on the locations and type of access.

The benchmark was configured to run both single-threaded and parallelized with OpenMP to take advantage of multiple memory controllers. The array of elements for the benchmark was set at 268,435,456 elements at compile time so that each array was 2048MB in size for a total memory footprint of approximately 6144 MB. The size was selected in line with the recommendation from STREAM that the array sizes be at least 4 times the total size of L3 cache available in the system. Pay special attention to the exact size of the L3 cache if V-Cache is present. An array-size offset was used so that the separate arrays for each parallelized thread would not share a Transparent Huge Page. The reason is that NUMA balancing may choose to migrate shared pages leading to some distortion of the results.

Table 3: Test workload: STREAM


gcc-11 (SUSE Linux) 11.2.1

Compiler flags

-Ofast -march=znver3 -mcmodel=medium -DOFFSET=512

OpenMP compiler flag


OpenMP environment variables



The march=znver3 is a reflection of the compiler available in SUSE Linux Enterprise 15 SP4 at the time of writing. It should be checked if a later GCC version is available in the Developer Tools Module that supported march=znver4. The number of openMP threads was selected to have at least one thread running for every memory channel by having one thread per L3 cache available. The OMP_PROC_BIND parameter spreads the threads such that one thread is bound to each available dedicated L3 cache to maximize available bandwidth. This can be verified using perf, as illustrated below with slight editing for formatting and clarity.

epyc:~ # perf record -e sched:sched_migrate_task ./stream
epyc:~ # perf script
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494780 prio=120 orig_cpu=0 dest_cpu=8
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494781 prio=120 orig_cpu=0 dest_cpu=16
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494782 prio=120 orig_cpu=0 dest_cpu=24
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494783 prio=120 orig_cpu=0 dest_cpu=32
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494784 prio=120 orig_cpu=0 dest_cpu=40
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494785 prio=120 orig_cpu=0 dest_cpu=48
          stream-nnn x: sched:sched_migrate_task: comm=stream pid=494786 prio=120 orig_cpu=0 dest_cpu=56

Several options were considered for the test system that were unnecessary for STREAM running on the AMD EPYC 9004 Series Processor but may be useful in other situations. STREAM performance can be limited if a load/store instruction stalls to fetch data. CPUs may automatically prefetch data based on historical behavior but it is not guaranteed. In limited cases, depending on the CPU and workload, this may be addressed by specifying -fprefetch-loop-arrays and depending on whether the workload is store-intensive, -mprefetchwt1. In the test system using AMD EPYC 9004 Series Processors, explicit prefetching did not help and was omitted. This is because an explicitly scheduled prefetch may disable a CPU's predictive algorithms and degrade performance. Similarly, for some workloads branch mispredictions can be a major problem, and in some cases breach mispredictions can be offset against I-Cache pressure by specifying -funroll-loops. In the case of STREAM on the test system, the CPU accurately predicted the branches rendering the unrolling of loops unnecessary. For math-intensive workloads it can be beneficial to link the application with -lmvec depending on the application. In the case of STREAM, the workload did not use significant math-based operations and so this option was not used. Some styles of code blocks and loops can also be optimized to use vectored operations by specifying -ftree-vectorize and explicitly adding support for CPU features such as -mavx2. In all cases, STREAM does not benefit as its operations are very basic. But it should be considered on an application-by-application basis and when building support libraries such as numerical libraries. In all cases, experimentation is recommended but caution advised. This holds particularly true when considering options like prefetch that may have been advisable on much older CPUs or completely different workloads. Such options are not universally beneficial or always suitable for modern CPUs such as the AMD EPYC 9004 Series Processors.

In the case of STREAM running on the AMD EPYC 9004 Series Processor, it was sufficient to enable -Ofast. This includes the -O3 optimizations to enable vectorization. In addition, it gives some leeway for optimizations that increase the code size with additional options for fast math that may not be standards-compliant.

For OpenMP, the SPREAD option was used to spread the load across L3 caches. OpenMP has a variety of different placement options if manually tuning placement. But careful attention should be paid to OMP_PLACES, given the importance of the L3 Cache topology in AMD EPYC 9004 Series Processors, if the operating system does not automatically place tasks appropriately. At the time of writing, it is not possible to specify l3cache as a place similar to what MPI has. In this case, the topology will need to be examined either with library support such as hwloc, directly via the sysfs or manually. While it is possible to guess via the CPUID, it is not recommended as CPUs may be offlined or the enumeration may vary between platforms because of BIOS implementations. An example specification of places based on L3 cache for the test system is:

{0:8,192:8},   {8:8,200:8},   {16:8,208:8},  {24:8,216:8},  {32:8,224:8},  {40:8,232:8},
{48:8,240:8},  {56:8,248:8},  {64:8,256:8},  {72:8,264:8},  {80:8,272:8},  {88:8,280:8},
{96:8,288:8},  {104:8,296:8}, {112:8,304:8}, {120:8,312:8}, {128:8,320:8}, {136:8,328:8},
{144:8,336:8}, {152:8,344:8}, {160:8,352:8}, {168:8,360:8}, {176:8,368:8}, {184:8,376:8}

Figure 2, “STREAM Bandwidth, Single Threaded and Parallelized” shows the reported bandwidth for the single and parallelized case. The single-threaded bandwidth for the basic Copy vector operation on a single core was 50.3 GB/sec. This is higher than the theoretical maximum of a single DIMM, but the IO die may interleave accesses, and caching effects and prefetch still apply. The total throughput for each parallel operation ranged from 1259 GB/sec to 1587 GB/sec depending on the type of operation and how efficiently memory bandwidth was used. This is very roughly scaling with the number of memory channels available on the machine with caching effects accounting for results exceeding theoretical maximums.

Note: STREAM scores

Higher STREAM scores can be reported by reducing the array sizes so that cache is partially used with the maximum score requiring that each threads memory footprint fits inside the L1 cache. Additionally, it is possible to achieve results closer to the theoretical maximum by manual optimization of the STREAM benchmark using vectored instructions and explicit scheduling of loads and stores. The purpose of this configuration was to illustrate the impact of properly binding a workload that can be fully parallelized with data-independent threads.

STREAM Bandwidth, Single Threaded and Parallelized
Figure 2: STREAM Bandwidth, Single Threaded and Parallelized

13.3 Test workload: NASA Parallel Benchmark

NASA Parallel Benchmark (NPB) is a small set of compute-intensive kernels designed to evaluate the performance of supercomputers. They are small compute kernels derived from Computational Fluid Dynamics (CFD) applications. The problem size can be adjusted for different memory sizes. Reference implementations exist for both MPI and OpenMP. This setup will focus on the MPI reference implementation.

While each application behaves differently, one common characteristic is that the workload is very context-switch intensive, barriers frequently yield the CPU to other tasks and the lifetime of individual processes can be very short-lived. The following paragraphs detail the tuning selected for this workload.

The most basic step is setting the CPU governor to performance although it is not mandatory. This can address issues with short-lived or mobile tasks failing to run long enough for a higher P-State to be selected even though the workload is very throughput-sensitive. The migration cost parameter is set to reduce the frequency in which the load balancer will move an individual task. The minimum granularity is adjusted to reduce overscheduling effects.

Depending on the computational kernel used, the workload may require a power-of-two number or a square number of processes to be used. However, note that using all available CPUs can mean that the application can contend with itself for CPU time. Furthermore, as IO may be issued to shared memory backed by disk, there are system threads that also need CPU time. Finally, if there is CPU contention, MPI tasks can be stalled waiting on an available CPU and OpenMPI may yield tasks prematurely if it detects there are more MPI tasks than CPUs available. These factors should be carefully considered when tuning for parallelized workloads in general and MPI workloads in particular.

In the specific case of testing NPB on the System Under Test, there was usually a limited advantage to limiting the number of CPUs used. For the Embarrassingly Parallel (EP) load in particular, it benefits from using all available CPUs. Hence, the default configuration used all available CPUs (384) which is both a power-of-two and square number of CPUs because it was a sensible starting point. However, this is not universally true. Using perf, it was found that some workloads are memory-bound and do not benefit from a high degree of parallelization. In addition, for the final configuration, some workloads were parallelized to have one task per L3 cache in the system to maximize cache usage. The exception was the Scalar Pentadiagonal (SP) workload which was both memory-bound and benefited from using all available CPUs. As the number of cores can vary between chips and the number of populated memory channels, the tuning parameters used for this test may not be universally true for all AMD EPYC platforms. This highlights that there is no universal good choice for optimizing a workload for a platform. Thus, experimentation and validation of tuning parameters is vital.

The basic compilation flags simply turned on all safe optimizations. The tuned flags used -Ofast which can be unsafe for some mathematical workloads but generated the correct output for NPB. The other flags used the optimal instructions available on the distribution compiler and vectorized some operations. GCC 11 is more strict in terms of matching types in Fortran. Depending on the version of NPB used, it may be necessary to specify the -fallow-argument-mismatch or -fallow-invalid-boz to compile unless the source code is modified.

As NPB uses shared files, an XFS partition was used for the temporary files. It is, however, only used for mapping shared files and is not a critical path for the benchmark and no IO tuning is necessary. In some cases, with MPI applications, it will be possible to use a tmpfs partition for OpenMPI. This avoids unnecessary IO assuming the increased physical memory usage does not cause the application to be paged out.

Table 4: Test workload: NASA Parallel Benchmark


gcc-11 (SUSE Linux) 11.2.1



Default compiler flags

-m64 -O3 -mcmodel=large

Default number processes


Selective number processes

bt=384 cg=192 ep=384 lu=192 mg=192 sp=192

Fortran compiler flags

-fallow-argument-mismatch -fallow-invalid-boz

Tuned compiler flags

-Ofast -march=znver3 -mtune=znver3 -ftree-vectorize

CPU governor performance

cpupower frequency-set -g performance

mpirun parameters

-mca btl ^openib,udapl -np 64 --bind-to l3cache:overload-allowed

mpirun environment


Figure 3, “NAS MPI Results” shows the time, as reported by the benchmark, for each of the kernels to complete.

NAS MPI Results
Figure 3: NAS MPI Results

The gcc-7-default test used the system compiler, all available CPUs, basic compilation options and the performance governor. The second test gcc-11-default used an alternative compiler. gcc-11-tuned used additional compilation options, and bound tasks to L3 caches gaining between 1.89% and 43.4% performance on average relative to gcc-7-default. Even taking into account that ep.D is an outlier because of the fact it is an embarrassingly parallel workload, the range of improvements is between 1.89% and 11.65%. The final test selective used processes that either used all CPUs, avoided heavy overloaded or limited processes to one per L3 cache, showing additional gains between 0.7% and 15.56% depending on the computational kernel.

In some cases, it will be necessary to compile an application that can run on different CPUs. In such cases, -march=znver3 may not be suitable if it generates binaries that are incompatible with other vendors. In such cases, it is possible to specify the ISA-specific options that are cross-compatible with many x86-based CPUs such as -mavx2, -mfma, -msse2 or msse4a while favoring optimal code generation for AMD EPYC 9004 Series Processors with -mtune=znver3. This can be used to strike a balance between excellent performance on a single CPU and great performance on multiple CPUs.

14 Tuning AMD EPYC 9004 Series Dense

As the AMD EPYC 9004 Series Classic and AMD EPYC Series Dense are ISA-compatible, no code tuning or compiler setting changes should be necessary. For Cloud environments, partitioning or any binding of Virtual CPUs to Physical CPUs may need to be adjusted to account for the increased number of cores. The additional cores may also allow additional containers or virtual machines to be hosted on the same physical machine without CPU contention. Similarly, the degree of parallelization for HPC workloads may need to be adjusted. In cases where the workload is tuned based on the number of CCD's, adjustments may be necessary for the increased number of CCDs. An exception are cases where the workload already hits scaling limits. While tuning based on the different number of CCXs is possible, it should only be necessary for applications with very strict latency requirements. As the size of the cache is halved, partitioning based on cache sizes may also need to be adjusted. In some cases, where workloads are tuned based on the output of tools like hwloc partitioning may adjust automatically but any static partitioning should be re-examined.

When configuring workloads for either the AMD EPYC 9004 Series Classic or the AMD EPYC 9004 Series Dense, the most important task is to set expectations. While super-linear scaling is possible, it should not be expected. It may be possible to achieve super-linear scaling in Cloud Environments for the number of instances hosted without performance loss if individual containers or virtual machines are not utilising 100% of CPU. However, it should be planned carefully and tested. This would be particularly true in cases where multiple instances are hosted that have different times of day or year for active phases. The normal expectation is a best case of 33% gain for CPU-intensive workloads due to the increased number of cores. But sub-linear scaling is common due to resource contention. Contention between SMT siblings, memory bandwidth, memory availability, memory interconnects, thread communication overhead or peripheral devices may prevent perfect linear scaling even for perfectly parallelized applications. Similarly, not all applications can scale perfectly. It is possible for performance to plateau and even regress as the degree of parallelization increases. Testing for parallelized workloads using NAS indicated that actual gains were between 4% and 21% for most workloads that did not have an inherent scaling limitation.

15 Conclusion

The introduction of the AMD EPYC 9004 Series Classic and AMD EPYC 9004 Series Dense Processors continues to push the boundaries of what is possible for memory and IO-bound workloads with much higher bandwidth and available number of channels. A properly configured and tuned workload can exceed the performance of many contemporary off-the-shelf solutions even when fully customized. The symmetric and balanced nature of the machine makes the task of tuning a workload considerably easier, given that each partition can have symmetric performance. And this is a property that can turn out particularly handy in virtualization, as each virtual machine can be assigned to each one of the partitions.

With SUSE Linux Enterprise 15 SP4, all the tools to monitor and tune a workload are readily available. You can extract the maximum performance and reliability running your applications on the 4th Generation AMD EPYC Processor platform.

16 Resources

For more information, refer to:

17 Legal notice

Copyright ©2006-2024 SUSE LLC and contributors. All rights reserved.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or (at your option) version 1.3; with the Invariant Section being this copyright notice and license. A copy of the license version 1.2 is included in the section entitled GNU Free Documentation License.

SUSE, the SUSE logo and YaST are registered trademarks of SUSE LLC in the United States and other countries. For SUSE trademarks, see http://www.suse.com/company/legal/. Linux is a registered trademark of Linus Torvalds. All other names or trademarks mentioned in this document may be trademarks or registered trademarks of their respective owners.

Documents published as part of the SUSE Best Practices series have been contributed voluntarily by SUSE employees and third parties. They are meant to serve as examples of how particular actions can be performed. They have been compiled with utmost attention to detail. However, this does not guarantee complete accuracy. SUSE cannot verify that actions described in these documents do what is claimed or whether actions described have unintended consequences. SUSE LLC, its affiliates, the authors, and the translators may not be held liable for possible errors or the consequences thereof.

Below we draw your attention to the license under which the articles are published.

18 GNU Free Documentation License

Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc. 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.

The purpose of this License is to make a manual, textbook, or other functional and useful document "free" in the sense of freedom: to assure everyone the effective freedom to copy and redistribute it, with or without modifying it, either commercially or non-commercially. Secondarily, this License preserves for the author and publisher a way to get credit for their work, while not being considered responsible for modifications made by others.

This License is a kind of "copyleft", which means that derivative works of the document must themselves be free in the same sense. It complements the GNU General Public License, which is a copyleft license designed for free software.

We have designed this License to use it for manuals for free software, because free software needs free documentation: a free program should come with manuals providing the same freedoms that the software does. But this License is not limited to software manuals; it can be used for any textual work, regardless of subject matter or whether it is published as a printed book. We recommend this License principally for works whose purpose is instruction or reference.

This License applies to any manual or other work, in any medium, that contains a notice placed by the copyright holder saying it can be distributed under the terms of this License. Such a notice grants a world-wide, royalty-free license, unlimited in duration, to use that work under the conditions stated herein. The "Document", below, refers to any such manual or work. Any member of the public is a licensee, and is addressed as "you". You accept the license if you copy, modify or distribute the work in a way requiring permission under copyright law.

A "Modified Version" of the Document means any work containing the Document or a portion of it, either copied verbatim, or with modifications and/or translated into another language.

A "Secondary Section" is a named appendix or a front-matter section of the Document that deals exclusively with the relationship of the publishers or authors of the Document to the Document's overall subject (or to related matters) and contains nothing that could fall directly within that overall subject. (Thus, if the Document is in part a textbook of mathematics, a Secondary Section may not explain any mathematics.) The relationship could be a matter of historical connection with the subject or with related matters, or of legal, commercial, philosophical, ethical or political position regarding them.

The "Invariant Sections" are certain Secondary Sections whose titles are designated, as being those of Invariant Sections, in the notice that says that the Document is released under this License. If a section does not fit the above definition of Secondary then it is not allowed to be designated as Invariant. The Document may contain zero Invariant Sections. If the Document does not identify any Invariant Sections then there are none.

The "Cover Texts" are certain short passages of text that are listed, as Front-Cover Texts or Back-Cover Texts, in the notice that says that the Document is released under this License. A Front-Cover Text may be at most 5 words, and a Back-Cover Text may be at most 25 words.

A "Transparent" copy of the Document means a machine-readable copy, represented in a format whose specification is available to the general public, that is suitable for revising the document straightforwardly with generic text editors or (for images composed of pixels) generic paint programs or (for drawings) some widely available drawing editor, and that is suitable for input to text formatters or for automatic translation to a variety of formats suitable for input to text formatters. A copy made in an otherwise Transparent file format whose markup, or absence of markup, has been arranged to thwart or discourage subsequent modification by readers is not Transparent. An image format is not Transparent if used for any substantial amount of text. A copy that is not "Transparent" is called "Opaque".

Examples of suitable formats for Transparent copies include plain ASCII without markup, Texinfo input format, LaTeX input format, SGML or XML using a publicly available DTD, and standard-conforming simple HTML, PostScript or PDF designed for human modification. Examples of transparent image formats include PNG, XCF and JPG. Opaque formats include proprietary formats that can be read and edited only by proprietary word processors, SGML or XML for which the DTD and/or processing tools are not generally available, and the machine-generated HTML, PostScript or PDF produced by some word processors for output purposes only.

The "Title Page" means, for a printed book, the title page itself, plus such following pages as are needed to hold, legibly, the material this License requires to appear in the title page. For works in formats which do not have any title page as such, "Title Page" means the text near the most prominent appearance of the work's title, preceding the beginning of the body of the text.

A section "Entitled XYZ" means a named subunit of the Document whose title either is precisely XYZ or contains XYZ in parentheses following text that translates XYZ in another language. (Here XYZ stands for a specific section name mentioned below, such as "Acknowledgements", "Dedications", "Endorsements", or "History".) To "Preserve the Title" of such a section when you modify the Document means that it remains a section "Entitled XYZ" according to this definition.

The Document may include Warranty Disclaimers next to the notice which states that this License applies to the Document. These Warranty Disclaimers are considered to be included by reference in this License, but only as regards disclaiming warranties: any other implication that these Warranty Disclaimers may have is void and has no effect on the meaning of this License.

You may copy and distribute the Document in any medium, either commercially or noncommercially, provided that this License, the copyright notices, and the license notice saying this License applies to the Document are reproduced in all copies, and that you add no other conditions whatsoever to those of this License. You may not use technical measures to obstruct or control the reading or further copying of the copies you make or distribute. However, you may accept compensation in exchange for copies. If you distribute a large enough number of copies you must also follow the conditions in section 3.

You may also lend copies, under the same conditions stated above, and you may publicly display copies.

If you publish printed copies (or copies in media that commonly have printed covers) of the Document, numbering more than 100, and the Document's license notice requires Cover Texts, you must enclose the copies in covers that carry, clearly and legibly, all these Cover Texts: Front-Cover Texts on the front cover, and Back-Cover Texts on the back cover. Both covers must also clearly and legibly identify you as the publisher of these copies. The front cover must present the full title with all words of the title equally prominent and visible. You may add other material on the covers in addition. Copying with changes limited to the covers, as long as they preserve the title of the Document and satisfy these conditions, can be treated as verbatim copying in other respects.

If the required texts for either cover are too voluminous to fit legibly, you should put the first ones listed (as many as fit reasonably) on the actual cover, and continue the rest onto adjacent pages.

If you publish or distribute Opaque copies of the Document numbering more than 100, you must either include a machine-readable Transparent copy along with each Opaque copy, or state in or with each Opaque copy a computer-network location from which the general network-using public has access to download using public-standard network protocols a complete Transparent copy of the Document, free of added material. If you use the latter option, you must take reasonably prudent steps, when you begin distribution of Opaque copies in quantity, to ensure that this Transparent copy will remain thus accessible at the stated location until at least one year after the last time you distribute an Opaque copy (directly or through your agents or retailers) of that edition to the public.

It is requested, but not required, that you contact the authors of the Document well before redistributing any large number of copies, to give them a chance to provide you with an updated version of the Document.

You may copy and distribute a Modified Version of the Document under the conditions of sections 2 and 3 above, provided that you release the Modified Version under precisely this License, with the Modified Version filling the role of the Document, thus licensing distribution and modification of the Modified Version to whoever possesses a copy of it. In addition, you must do these things in the Modified Version:

  1. Use in the Title Page (and on the covers, if any) a title distinct from that of the Document, and from those of previous versions (which should, if there were any, be listed in the History section of the Document). You may use the same title as a previous version if the original publisher of that version gives permission.

  2. List on the Title Page, as authors, one or more persons or entities responsible for authorship of the modifications in the Modified Version, together with at least five of the principal authors of the Document (all of its principal authors, if it has fewer than five), unless they release you from this requirement.

  3. State on the Title page the name of the publisher of the Modified Version, as the publisher.

  4. Preserve all the copyright notices of the Document.

  5. Add an appropriate copyright notice for your modifications adjacent to the other copyright notices.

  6. Include, immediately after the copyright notices, a license notice giving the public permission to use the Modified Version under the terms of this License, in the form shown in the Addendum below.

  7. Preserve in that license notice the full lists of Invariant Sections and required Cover Texts given in the Document's license notice.

  8. Include an unaltered copy of this License.

  9. Preserve the section Entitled "History", Preserve its Title, and add to it an item stating at least the title, year, new authors, and publisher of the Modified Version as given on the Title Page. If there is no section Entitled "History" in the Document, create one stating the title, year, authors, and publisher of the Document as given on its Title Page, then add an item describing the Modified Version as stated in the previous sentence.

  10. Preserve the network location, if any, given in the Document for public access to a Transparent copy of the Document, and likewise the network locations given in the Document for previous versions it was based on. These may be placed in the "History" section. You may omit a network location for a work that was published at least four years before the Document itself, or if the original publisher of the version it refers to gives permission.

  11. For any section Entitled "Acknowledgements" or "Dedications", Preserve the Title of the section, and preserve in the section all the substance and tone of each of the contributor acknowledgements and/or dedications given therein.

  12. Preserve all the Invariant Sections of the Document, unaltered in their text and in their titles. Section numbers or the equivalent are not considered part of the section titles.

  13. Delete any section Entitled "Endorsements". Such a section may not be included in the Modified Version.

  14. Do not retitle any existing section to be Entitled "Endorsements" or to conflict in title with any Invariant Section.

  15. Preserve any Warranty Disclaimers.

If the Modified Version includes new front-matter sections or appendices that qualify as Secondary Sections and contain no material copied from the Document, you may at your option designate some or all of these sections as invariant. To do this, add their titles to the list of Invariant Sections in the Modified Version's license notice. These titles must be distinct from any other section titles.

You may add a section Entitled "Endorsements", provided it contains nothing but endorsements of your Modified Version by various parties--for example, statements of peer review or that the text has been approved by an organization as the authoritative definition of a standard.

You may add a passage of up to five words as a Front-Cover Text, and a passage of up to 25 words as a Back-Cover Text, to the end of the list of Cover Texts in the Modified Version. Only one passage of Front-Cover Text and one of Back-Cover Text may be added by (or through arrangements made by) any one entity. If the Document already includes a cover text for the same cover, previously added by you or by arrangement made by the same entity you are acting on behalf of, you may not add another; but you may replace the old one, on explicit permission from the previous publisher that added the old one.

The author(s) and publisher(s) of the Document do not by this License give permission to use their names for publicity for or to assert or imply endorsement of any Modified Version.

You may combine the Document with other documents released under this License, under the terms defined in section 4 above for modified versions, provided that you include in the combination all of the Invariant Sections of all of the original documents, unmodified, and list them all as Invariant Sections of your combined work in its license notice, and that you preserve all their Warranty Disclaimers.

The combined work need only contain one copy of this License, and multiple identical Invariant Sections may be replaced with a single copy. If there are multiple Invariant Sections with the same name but different contents, make the title of each such section unique by adding at the end of it, in parentheses, the name of the original author or publisher of that section if known, or else a unique number. Make the same adjustment to the section titles in the list of Invariant Sections in the license notice of the combined work.

In the combination, you must combine any sections Entitled "History" in the various original documents, forming one section Entitled "History"; likewise combine any sections Entitled "Acknowledgements", and any sections Entitled "Dedications". You must delete all sections Entitled "Endorsements".

You may make a collection consisting of the Document and other documents released under this License, and replace the individual copies of this License in the various documents with a single copy that is included in the collection, provided that you follow the rules of this License for verbatim copying of each of the documents in all other respects.

You may extract a single document from such a collection, and distribute it individually under this License, provided you insert a copy of this License into the extracted document, and follow this License in all other respects regarding verbatim copying of that document.

A compilation of the Document or its derivatives with other separate and independent documents or works, in or on a volume of a storage or distribution medium, is called an "aggregate" if the copyright resulting from the compilation is not used to limit the legal rights of the compilation's users beyond what the individual works permit. When the Document is included in an aggregate, this License does not apply to the other works in the aggregate which are not themselves derivative works of the Document.

If the Cover Text requirement of section 3 is applicable to these copies of the Document, then if the Document is less than one half of the entire aggregate, the Document's Cover Texts may be placed on covers that bracket the Document within the aggregate, or the electronic equivalent of covers if the Document is in electronic form. Otherwise they must appear on printed covers that bracket the whole aggregate.

Translation is considered a kind of modification, so you may distribute translations of the Document under the terms of section 4. Replacing Invariant Sections with translations requires special permission from their copyright holders, but you may include translations of some or all Invariant Sections in addition to the original versions of these Invariant Sections. You may include a translation of this License, and all the license notices in the Document, and any Warranty Disclaimers, provided that you also include the original English version of this License and the original versions of those notices and disclaimers. In case of a disagreement between the translation and the original version of this License or a notice or disclaimer, the original version will prevail.

If a section in the Document is Entitled "Acknowledgements", "Dedications", or "History", the requirement (section 4) to Preserve its Title (section 1) will typically require changing the actual title.

You may not copy, modify, sublicense, or distribute the Document except as expressly provided for under this License. Any other attempt to copy, modify, sublicense or distribute the Document is void, and will automatically terminate your rights under this License. However, parties who have received copies, or rights, from you under this License will not have their licenses terminated so long as such parties remain in full compliance.

The Free Software Foundation may publish new, revised versions of the GNU Free Documentation License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. See http://www.gnu.org/copyleft/.

Each version of the License is given a distinguishing version number. If the Document specifies that a particular numbered version of this License "or any later version" applies to it, you have the option of following the terms and conditions either of that specified version or of any later version that has been published (not as a draft) by the Free Software Foundation. If the Document does not specify a version number of this License, you may choose any version ever published (not as a draft) by the Free Software Foundation.

Copyright (c) YEAR YOUR NAME.
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.2
or any later version published by the Free Software Foundation;
with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
A copy of the license is included in the section entitled "GNU
 Free Documentation License".

If you have Invariant Sections, Front-Cover Texts and Back-Cover Texts, replace the "with...Texts". line with this:

with the Invariant Sections being LIST THEIR TITLES, with the
Front-Cover Texts being LIST, and with the Back-Cover Texts being LIST.

If you have Invariant Sections without Cover Texts, or some other combination of the three, merge those two alternatives to suit the situation.

If your document contains nontrivial examples of program code, we recommend releasing these examples in parallel under your choice of free software license, such as the GNU General Public License, to permit their use in free software.