AmpereOne Performance On Linux 6.11 Kernel, 4K vs. 64K Page Size Comparison

[phoronix]

Phoronix: AmpereOne Performance On Linux 6.11 Kernel, 4K vs. 64K Page Size Comparison

Continuing on with the AmpereOne performance benchmarking while having the AmpereOne A192-32X in the lab within a Supermicro ARS-211M-NR R13SPD server, the next set of benchmarks is looking at the performance when using the near-final Linux 6.11 kernel. Additionally, quantifying the performance impact of using the ARM64 64K page size kernel as an alternative to the default 4K page size.

https://www.phoronix.com/review/ampereone-64k-linux611

[jeisom]

Does this hardware support 16K sizes like Apple's hardware? If so, I wonder how that would do. The asahi builds are using 16K by default and on AS macs there was a noticeable improvement in performance in some workloads if I recall correctly.

[V1tol]

I wonder if it is possible to use 16k or 64k size on x86-64 and if yes - what's the performance difference will be.

[JoeSpeed]

Does this hardware support 16K sizes like Apple's hardware? …

Yes, 16K page is supported. Performance lands between 4K and 64K as you would expect.

[espi]

I wonder if it is possible to use 16k or 64k size on x86-64 and if yes - what's the performance difference will be.

x86-64 supports only 4kb pages with 2MB/1GB huge pages.

I sometimes wonder if its possible to enforce all huge pages. Memory usage would be awful though.

[Mitch]

I imagine going too high on page sizes could swing performance back if it causes more cache misses. L1 cache is often pretty small and I think even smaller on efficiency-types of cores

[andyhhp]

x86-64 supports only 4kb pages with 2MB/1GB huge pages.

Architecturally, yes. Microarchitecturally, AMD Zen2 and later will create 8k/16k/32k TLB entries if 2/4/8 adjacent PTEs are compatible (same perms, adjacent addresses); which they frequently are for libraries and binaries in memory.

[caligula]

I sometimes wonder if its possible to enforce all huge pages. Memory usage would be awful though.

Yes there are multiple options in kernelconfig. Unfortunately the force option increases usage quite a bit. 16-64 kB would be a lot better. 4 kB is too small when typical desktop systems already support up to 128-256 GB. I've had 64 GB for a year or two. Firefox easily consumes 16+ GB. Chromium even shows the usage per tab. Usually at least 70-80 MB per tab for small pages. 100 tabs -> 8 GB. Long uptime -> more wasted gigs.

[sarmad]

Wow, that's a bigger boost than I thought it would be.

Now we need a test for the memory overhead.

[blackshard]

I wonder where the 4K overhead comes from and how this interacts with libc malloc arena sizes; perhaps some overheader for the 4K pages can be reduced with some different heuristics in libc kernel pages allocations.