AMD Instinct MI300X system optimization

AMD Instinct MI300X system optimization#

2024-12-16

21 min read time

Applies to Linux

This document covers essential system settings and management practices required to configure your system effectively. Ensuring that your system operates correctly is the first step before delving into advanced performance tuning.

The main topics of discussion in this document are:

System settings
System management

System settings#

This guide discusses system settings that are required to configure your system for AMD Instinct™ MI300X accelerators. It is important to ensure a system is functioning correctly before trying to improve its overall performance. In this section, the settings discussed mostly ensure proper functionality of your Instinct-based system. Some settings discussed are known to improve performance for most applications running on a MI300X system. See AMD Instinct MI300X workload optimization for how to improve performance for specific applications or workloads.

System BIOS settings#

AMD EPYC 9004-based systems#

For maximum MI300X GPU performance on systems with AMD EPYC™ 9004-series processors and AMI System BIOS, the following configuration of system BIOS settings has been validated. These settings must be used for the qualification process and should be set as default values in the system BIOS. Analogous settings for other non-AMI System BIOS providers could be set similarly. For systems with Intel processors, some settings may not apply or be available as listed in the following table.

Each row in the table details a setting but the specific location within the BIOS setup menus may be different, or the option may not be present.

BIOS setting location	Parameter	Value	Comments
Advanced / PCI subsystem settings	Above 4G decoding	Enabled	GPU large BAR support.
Advanced / PCI subsystem settings	SR-IOV support	Enabled	Enable single root IO virtualization.
AMD CBS / GPU common options	Global C-state control	Auto	Global C-states – do not disable this menu item).
AMD CBS / GPU common options	CCD/Core/Thread enablement	Accept	May be necessary to enable the SMT control menu.
AMD CBS / GPU common options / performance	SMT control	Disable	Set to Auto if the primary application is not compute-bound.
AMD CBS / DF common options / memory addressing	NUMA nodes per socket	Auto	Auto = NPS1. At this time, the other options for NUMA nodes per socket should not be used.
AMD CBS / DF common options / memory addressing	Memory interleaving	Auto	Depends on NUMA nodes (NPS) setting.
AMD CBS / DF common options / link	4-link xGMI max speed	32 Gbps	Auto results in the speed being set to the lower of the max speed the motherboard is designed to support and the max speed of the CPU in use.
AMD CBS / NBIO common options	IOMMU	Enabled
AMD CBS / NBIO common options	PCIe ten bit tag support	Auto
AMD CBS / NBIO common options / SMU common options	Determinism control	Manual
AMD CBS / NBIO common options / SMU common options	Determinism slider	Power
AMD CBS / NBIO common options / SMU common options	cTDP control	Manual	Set cTDP to the maximum supported by the installed CPU.
AMD CBS / NBIO common options / SMU common options	cTDP	400	Value in watts.
AMD CBS / NBIO common options / SMU common options	Package power limit control	Manual	Set package power limit to the maximum supported by the installed CPU.
AMD CBS / NBIO common options / SMU common options	Package power limit	400	Value in watts.
AMD CBS / NBIO common options / SMU common options	xGMI link width control	Manual	Set package power limit to the maximum supported by the installed CPU.
AMD CBS / NBIO common options / SMU common options	xGMI force width control	Force
AMD CBS / NBIO common options / SMU common options	xGMI force link width	2	0: Force xGMI link width to x2 1: Force xGMI link width to x8 2: Force xGMI link width to x16
AMD CBS / NBIO common options / SMU common options	xGMI max speed	Auto	Auto results in the speed being set to the lower of the max speed the motherboard is designed to support and the max speed of the CPU in use.
AMD CBS / NBIO common options / SMU common options	APBDIS	1	Disable DF (data fabric) P-states
AMD CBS / NBIO common options / SMU common options	DF C-states	Auto
AMD CBS / NBIO common options / SMU common options	Fixed SOC P-state	P0
AMD CBS / security	TSME	Disabled	Memory encryption

GRUB settings#

In any modern Linux distribution, the /etc/default/grub file is used to configure GRUB. In this file, the string assigned to GRUB_CMDLINE_LINUX is the command line parameters that Linux uses during boot.

Appending strings via Linux command line#

It is recommended to append the following strings in GRUB_CMDLINE_LINUX.

pci=realloc=off: With this setting Linux is able to unambiguously detect all GPUs of the MI300X-based system because this setting disables the automatic reallocation of PCI resources. It’s used when Single Root I/O Virtualization (SR-IOV) Base Address Registers (BARs) have not been allocated by the BIOS. This can help avoid potential issues with certain hardware configurations.
iommu=pt: The iommu=pt setting enables IOMMU pass-through mode. When in pass-through mode, the adapter does not need to use DMA translation to the memory, which can improve performance.

IOMMU is a system specific IO mapping mechanism and can be used for DMA mapping and isolation. This can be beneficial for virtualization and device assignment to virtual machines. It is recommended to enable IOMMU support.

For a system that has AMD host CPUs add this to GRUB_CMDLINE_LINUX:

iommu=pt

Otherwise, if the system has Intel host CPUs add this instead to GRUB_CMDLINE_LINUX:

intel_iommu=on iommu=pt

Update GRUB#

Update GRUB to use the modified configuration:

sudo grub2-mkconfig -o /boot/grub2/grub.cfg

On some Debian systems, the grub2-mkconfig command may not be available. Instead, check for the presence of grub-mkconfig. Additionally, verify that you have the correct version by using the following command:

grub-mkconfig -version

Operating system settings#

CPU core states (C-states)#

There are several core states (C-states) that an AMD EPYC CPU can idle within:

C0: active. This is the active state while running an application.
C1: idle. This state consumes less power compared to C0, but can quickly return to the active state (C0) with minimal latency.
C2: idle and power-gated. This is a deeper sleep state and will have greater latency when moving back to the active (C0) state as compared to when the CPU is coming out of C1.

Disabling C2 is important for running with a high performance, low-latency network. To disable the C2 state, install the cpupower tool using your Linux distribution’s package manager. cpupower is not a base package in most Linux distributions. The specific package to be installed varies per Linux distribution.

Ubuntu

sudo apt install linux-tools-common

RHEL

sudo yum install cpupowerutils

SLES

sudo zypper install cpupower

Now, to disable power-gating on all cores run the following on Linux systems, run the following command.

cpupower idle-set -d 2

/proc and /sys file system settings#

Disable NUMA auto-balancing#

The NUMA balancing feature allows the OS to scan memory and attempt to migrate to a DIMM that is logically closer to the cores accessing it. This causes an overhead because the OS is second-guessing your NUMA allocations but may be useful if the NUMA locality access is very poor. Applications can therefore, in general, benefit from disabling NUMA balancing; however, there are workloads where doing so is detrimental to performance. Test this setting by toggling the numa_balancing value and running the application; compare the performance of one run with this set to 0 and another run with this to 1.

Run the command cat /proc/sys/kernel/numa_balancing to check the current NUMA (Non-Uniform Memory Access) settings. Output 0 indicates this setting is disabled. If no output or output is 1, run the command sudo sh -c \\'echo 0 > /proc/sys/kernel/numa_balancing to disable it.

For these settings, the env_check.sh script automates setting, resetting, and checking your environments. Find the script at ROCm/triton.

Run the script as follows to set or reset the settings:

./env_check.sh [set/reset/check]

Tip

Use ./env_check.sh -h for help info.

Automate disabling NUMA auto-balance using Cron#

The Disable NUMA auto-balancing section describes the command to disable NUMA auto-balance. To automate the command with Cron, edit the crontab configuration file for the root user:

sudo crontab -e

Add the following Cron entry to run the script at a specific interval:
```
@reboot sh -c 'echo 0 > /proc/sys/kernel/numa_balancing'
```
Save the file and exit the text editor.
Optionally, restart the system to apply changes by issuing sudo reboot.
Verify your new configuration.
```
cat /proc/sys/kernel/numa_balancing
```
The /proc/sys/kernel/numa_balancing file controls NUMA balancing in the Linux kernel. If the value in this file is set to 0, the NUMA balancing is disabled. If the value is set to 1, NUMA balancing is enabled.

Note

Disabling NUMA balancing should be done cautiously and for specific reasons, such as performance optimization or addressing particular issues. Always test the impact of disabling NUMA balancing in a controlled environment before applying changes to a production system.

Environment variables#

HIP provides an environment variable export HIP_FORCE_DEV_KERNARG=1 that can put arguments of HIP kernels directly to device memory to reduce the latency of accessing those kernel arguments. It can improve performance by 2 to 3 µs for some kernels.

It is recommended to set the following environment variable:

export HIP_FORCE_DEV_KERNARG=1

Note

This is the default option as of ROCm 6.2.

Change affinity of ROCm helper threads#

This change prevents internal ROCm threads from having their CPU core affinity mask set to all CPU cores available. With this setting, the threads inherit their parent’s CPU core affinity mask. If you have any questions regarding this setting, contact your MI300X platform vendor. To enable this setting, enter the following command:

export HSA_OVERRIDE_CPU_AFFINITY_DEBUG=0

IOMMU configuration – systems with 256 CPU threads#

For systems that have 256 logical CPU cores or more, setting the input-output memory management unit (IOMMU) configuration to disabled can limit the number of available logical cores to 255. The reason is that the Linux kernel disables X2APIC in this case and falls back to Advanced Programmable Interrupt Controller (APIC), which can only enumerate a maximum of 255 (logical) cores.

If SMT is enabled by setting CCD/Core/Thread Enablement > SMT Control to enable, you can apply the following steps to the system to enable all (logical) cores of the system:

In the server BIOS, set IOMMU to Enabled.
When configuring the GRUB boot loader, add the following argument for the Linux kernel: iommu=pt.
Update GRUB.
Reboot the system.
Verify IOMMU passthrough mode by inspecting the kernel log via dmesg:
```
dmesg | grep iommu
```

[...]
[   0.000000] Kernel command line: [...] iommu=pt
[...]

Once the system is properly configured, ROCm software can be installed.

System management#

To optimize system performance, it’s essential to first understand the existing system configuration parameters and settings. ROCm offers several CLI tools that can provide system-level information, offering valuable insights for optimizing user applications.

For a complete guide on how to install, manage, or uninstall ROCm on Linux, refer to Quick start installation guide. For verifying that the installation was successful, refer to the Post-installation instructions. Should verification fail, consult System debugging.

Hardware verification with ROCm#

The ROCm platform provides tools to query the system structure. These include ROCm SMI and ROCm Bandwidth Test.

ROCm SMI#

To query your GPU hardware, use the rocm-smi command. ROCm SMI lists GPUs available to your system – with their device ID and their respective firmware (or VBIOS) versions.

The following screenshot shows that all 8 GPUs of MI300X are recognized by ROCm. Performance of an application could be otherwise suboptimal if, for example, out of the 8 GPUs only 5 of them are recognized.

To see the system structure, the localization of the GPUs in the system, and the fabric connections between the system components, use the command rocm-smi --showtopo.

The first block of the output shows the distance between the GPUs similar to what the numactl command outputs for the NUMA domains of a system. The weight is a qualitative measure for the “distance” data must travel to reach one GPU from another one. While the values do not carry a special, or “physical” meaning, the higher the value the more hops are needed to reach the destination from the source GPU. This information has performance implication for a GPU-based application that moves data among GPUs. You can choose a minimum distance among GPUs to be used to make the application more performant.

The second block has a matrix named Hops between two GPUs, where:

1 means the two GPUs are directly connected with xGMI,
2 means both GPUs are linked to the same CPU socket and GPU communications will go through the CPU, and
3 means both GPUs are linked to different CPU sockets so communications will go through both CPU sockets. This number is one for all GPUs in this case since they are all connected to each other through the Infinity Fabric links.

The third block outputs the link types between the GPUs. This can either be XGMI for AMD Infinity Fabric links or PCIE for PCIe Gen5 links.

The fourth block reveals the localization of a GPU with respect to the NUMA organization of the shared memory of the AMD EPYC processors.

To query the compute capabilities of the GPU devices, use rocminfo command. It lists specific details about the GPU devices, including but not limited to the number of compute units, width of the SIMD pipelines, memory information, and instruction set architecture (ISA). The following is the truncated output of the command:

For a complete list of architecture (such as CDNA3) and LLVM target names (such gfx942 for MI300X), refer to the Supported GPUs section of the System requirements for Linux page.

Deterministic clock#

Use the command rocm-smi --setperfdeterminism 1900 to set the max clock speed up to 1900 MHz instead of the default 2100 MHz. This can reduce the chance of a PCC event lowering the attainable GPU clocks. This setting will not be required for new IFWI releases with the production PRC feature. Restore this setting to its default value with the rocm-smi -r command.

ROCm Bandwidth Test#

The section Hardware verification with ROCm showed how the command rocm-smi --showtopo can be used to view the system structure and how the GPUs are connected. For more details on the link bandwidth, rocm-bandwidth-test can run benchmarks to show the effective link bandwidth between the components of the system.

You can install ROCm Bandwidth Test, which can test inter-device bandwidth, using the following package manager commands:

Ubuntu

sudo apt install rocm-bandwidth-test

RHEL

sudo yum install rocm-bandwidth-test

SLES

sudo zypper install rocm-bandwidth-test

Alternatively, you can download the source code from ROCm/rocm_bandwidth_test and build from source.

The output will list the available compute devices (CPUs and GPUs), including their device ID and PCIe ID. The following screenshot is an example of the beginning part of the output of running rocm-bandwidth-test. It shows the devices present in the system.

The output will also show a matrix that contains a 1 if a device can communicate to another device (CPU and GPU) of the system and it will show the NUMA distance – similar to rocm-smi.

Inter-device distance:

Inter-device NUMA distance:

The output also contains the measured bandwidth for unidirectional and bidirectional transfers between the devices (CPU and GPU):

Unidirectional bandwidth:

Bidirectional bandwidth

Abbreviations#

AMI: American Megatrends International
APBDIS: Algorithmic Performance Boost Disable
ATS: Address Translation Services
BAR: Base Address Register
BIOS: Basic Input/Output System
CBS: Common BIOS Settings
CLI: Command Line Interface
CPU: Central Processing Unit
cTDP: Configurable Thermal Design Power
DDR5: Double Data Rate 5 DRAM
DF: Data Fabric
DIMM: Dual In-line Memory Module
DMA: Direct Memory Access
DPM: Dynamic Power Management
GPU: Graphics Processing Unit
GRUB: Grand Unified Bootloader
HPC: High Performance Computing
IOMMU: Input-Output Memory Management Unit
ISA: Instruction Set Architecture
LCLK: Link Clock Frequency
NBIO: North Bridge Input/Output
NUMA: Non-Uniform Memory Access
PCC: Power Consumption Control
PCI: Peripheral Component Interconnect
PCIe: PCI Express
POR: Power-On Reset
SIMD: Single Instruction, Multiple Data
SMT: Simultaneous Multi-threading
SMI: System Management Interface
SOC: System On Chip
SR-IOV: Single Root I/O Virtualization
TP: Tensor Parallelism
TSME: Transparent Secure Memory Encryption
X2APIC: Extended Advanced Programmable Interrupt Controller
xGMI: Inter-chip Global Memory Interconnect