.jpg){kind=link}
AMD is working on a new technology called "Process Adaptive autoNUMA" or PAN for short that is showing some promising early performance numbers as per internal AMD testing. Fellow website Phoronix spotted the development in a request for comment (RFC) patch sent in by AMD"s Bharat B Rao that talks about this newly proposed PAN feature.
The Process Adaptive autoNUMA or PAN, as the name suggests, moves from a static auto-scan to an adaptive approach by collecting scan rates based on a per-process level instead of the currently used per-thread stats. The patch note suggests that by doing so, the new algorithm is better able to understand various application behaviors under NUMA or Non-Uniform Memory Access architectures.
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.
Under NUMA, a multi-processor system like an AMD EPYC server can access its remote memory as well. The NUMA design helps to add scalability and bandwidth and the new PAN patch can further improve on that.
Testing conducted on a dual-socket (2P) AMD Milan 64 core/128 thread system shows promising gains in some of the workloads. For example, there is nearly a 15% improvement in the Graph500 supercomputing benchmark as the PAN-enabled system was 14.93% better. There is also close to a 10 % gain in two out of four NAS Parallel benchmarks. You can find the full test results in the RFC patch linked at the source below.