Intel's Rewrite Of Linux MM CID Code Showing Some Nice Gains For AMD
Posted last month were new Linux kernel scheduler-related patches rewriting the MM CID management code. The main takeaway for end-users from this set of 19 Linux kernel patches from an Intel engineer was seeing 14~18% improvement in a PostgreSQL database benchmark but that more benchmarks were needed. Curiosity got the best of me and I recently tested these patches on an AMD EPYC server to seeing some very enticing results for this in-development code.
Those interested in all the low-level technical details of this MM CID management rewrite can see this patch series from mid-October by Intel Fellow Thomas Gleixner. With the call-out for additional benchmarks from that patch series, I ran some benchmarks on an AMD EPYC test server for seeing the performance impact in some common server workloads.
For my testing I was running against Gleixner's tglx/devel.git rseq/cid Git branch that contains his rseq/perf patches plus the MM CID code rewrite patches. He did post a new revision of the patches this week to that branch but my testing was done in October against his original patch series.
I compared the "rseq/cid" Git branch with these patches against a mainline Linux 6.19-rc1 Git kernel for which that branch was originally based. The same kernel configuration and compiler were used for both kernel builds.
Using a dual AMD EPYC 9965 server (combined 384 cores / 768 threads) and running Ubuntu 25.10 with these custom kernel builds tested, I ran dozens of different benchmarks to see the impact of the rseq/cid kernel patches compared to the upstream Linux 6.18 state. Benchmarks on additional hardware and of this week's very newest rseq/cid patches may come if there is sufficient Phoronix reader interest to justify the testing time needed.