Skip to content

[Benchmarks] Update perf readme #19925

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 2, 2025
+11 −27
Merged

[Benchmarks] Update perf readme #19925

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 11 additions & 27 deletions devops/scripts/benchmarks/PERFORMANCE_TUNING.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,38 +6,16 @@ For framework-specific information, see [README.md](README.md) and [CONTRIB.md](
## Table of Contents

- [Overview](#overview)
- [System Configuration](#system-configuration)
- [CPU Tuning](#cpu-tuning)
- [GPU Configuration](#gpu-configuration)
- [Driver Version](#driver-version)
- [Perf Configuration](#perf-configuration)
- [Environment Variables](#environment-variables)

## Overview

Performance benchmarking requires a stable and optimized system environment to produce reliable and reproducible results. This guide covers essential system tuning steps for reducing run-to-run variance in benchmark results.

## System Configuration

### Kernel Parameters

Add the following to `/etc/default/grub` in `GRUB_CMDLINE_LINUX`:
```
# Disable CPU frequency scaling
# intel_pstate=disable

# Isolate CPUs for benchmark workloads (example: reserve cores 2-7), preventing other processes
# from using them.
# isolcpus=2-7

GRUB_CMDLINE_LINUX="intel_pstate=disable isolcpus=2-7 <other_options>"
```

Update GRUB and reboot:
```bash
sudo update-grub
sudo reboot
```

## CPU Tuning

### CPU Frequency Scaling
Expand Down Expand Up @@ -66,7 +44,7 @@ After=multi-user.target

[Service]
Type=oneshot
ExecStart=/usr/bin/cpupower frequency-set --governor performance && sysctl --system
ExecStart=/usr/bin/cpupower frequency-set --governor performance

[Install]
WantedBy=multi-user.target
Expand Down Expand Up @@ -99,7 +77,12 @@ cat /sys/class/drm/card1/device/vendor # Should be 0x8086 for Intel
cat /sys/class/drm/card1/device/device # Device ID
```

Verify the max frequency is set to the true max. For Arc B580, the maximum frequency is 2850 MHz. To see this value, run “cat /sys/class/drm/card1/device/tile0/gt0/freq0/max_freq”. If the above value is not equal to the max frequency, set it as such:
Verify the max frequency is set to the true max. For Arc B580, the maximum frequency is 2850 MHz. To see this value, run:
```bash
cat /sys/class/drm/card1/device/tile0/gt0/freq0/max_freq
```

If the above value is not equal to the max frequency, set it as such:
```bash
# Arc B580 (Battlemage)
echo 2850 > /sys/class/drm/card1/device/tile0/gt0/freq0/max_freq
Expand All @@ -118,9 +101,9 @@ max_freq=$(cat /sys/class/drm/card1/gt_max_freq_mhz)
echo $max_freq | sudo tee /sys/class/drm/card1/gt_min_freq_mhz
```

The result can be verified using tools such as oneprof or unitrace to track frequency over time for some arbitrary benchmark (many iterations of a small problem size is recommended). The frequency should remain fixed assuming thermal throttling does not occur.
The result can be verified using tools such as `oneprof` or `unitrace` to track frequency over time for some arbitrary benchmark (many iterations of a small problem size is recommended). The frequency should remain fixed assuming thermal throttling does not occur.

## Driver version
## Driver Version
Make sure you are using the latest driver (Ubuntu)
```bash
sudo apt update && sudo apt upgrade
Expand All @@ -145,6 +128,7 @@ Make the setting persistent across reboots by adding it to sysctl configuration:
echo 'kernel.perf_event_paranoid = -1' | sudo tee -a /etc/sysctl.d/99-perf.conf

# Apply immediately
sudo sysctl kernel.perf_event_paranoid=-1
sudo sysctl -p
```

Expand Down