AMD offers new adaptive algorithm, Process Adaptive autoNUMA, to increase performance in Linux

Jason R. Wilson
Deci Posts MLPerf Benchmarks of New NLP Model, Achieves A Massive 6.46x Gain With AMD EPYC Milan-X CPUs 1

AMD's open-source department recently revealed a new kernel feature called Process Adaptive autoNUMA (PAN). The new feature allows for increased performance in specific server workloads in Linux operating systems.

AMD 'panning' for results with their newest Process Adaptive autoNUMA feature, boosting performance in specific server processes in Linux OS

PAN, or Process Adaptive autoNUMA, is a constantly adjusting algorithm that computes the AutoNUMA scan duration. Utilizing the PAN Linux kernel build, AMD open-source engineers found that the Graph500 interconnect HPC benchmark profits from a higher performance by up to 14.93% compared to a standard Linux kernel.

Related Story Intel Outlines 40 TOPS NPU Performance As Minimum Requirement For Windows Copilot & AI PC Platforms

The network-attached storage benchmark test displays an improvement of eight percent. PageRank and other various results only demonstrated less than 1% in all other trials, with PageRank only showing an increase of 0.37 percent. It is to note that there are only a handful of tests assessed by AMD engineers at this time. If the new kernel series by AMD for Linux succeeds past the RFC stages, other kernel programmers can use the PAN programming and be programmed into the final kernel.

In this new approach (Process Adaptive autoNUMA or PAN), we gather NUMA fault stats at per-process level which allows for capturing the application behaviour better. In addition, the algorithm learns and adjusts the scan rate based on remote fault rate. By not sticking to a static threshold, the algorithm can respond better to different workload behaviours. Since the threads of a processes are already considered as a group, we add a bunch of metrics to the task's [memory management] to track the various types of faults and derive the scan rate from them. The new per-process fault stats contribute only to the per-process scan period calculation, while the existing per-thread stats continue to contribute towards the numa_group stats which eventually determine the thresholds for migrating memory and threads across nodes.

— an explanation by Bharata B. Rao of AMD in reference to the request for comments (RFC) of their Linux kernel patch series

via Phoronix

Only about 400 lines of new code have appeared to enhance the Linux NUMA behavior. No Linux kernel developers have currently shown interest in implementing the Process Adaptive autoNUMA submission. However, suppose there are developers interested in the AMD PAN. In that case, they are to direct themselves to the RFC series for information and the ability to use the feature before its full release.

News Source: Phoronix

Share this story

Deal of the Day

Comments