Based on Richards comments and the unexpected presence of SSE and MMX in the benchmarks, I have added two new benchmark types.
fullbenchK = -O2 with no SSE and no MMX in the decompress or compress routines. (Both LZ4 and LZJB have MMX and SSE disabled) fullbenchK3 = -O3 with no SSE and no MMX in the decompress or compress routines. (Both LZ4 and LZJB have MMX and SSE disabled) I verified there was no mmx or sse in these using objdump. The original version used "movdqu" occasionally. The other "fullbench" programs do not tune down optimizations, so depending on core they may or may not have SSE or MMX. I keep them around for comparative testing. Testing on the "mozilla" file from silesia shows a performance drop in my new method of less than 1% at both O3 and O2 without SSE or MMX on my AND FX 8150. Difference in speed between Stock and my experimental version is 159% at -O2 with no sse or mmx. The BSD version is slower at all tests except for -O2 WITH sse and mmx. Full data from that test run: "Kernel SSE/MMX de-optimisation tests",,,,,, "AMD FX 8150 – 3.6Ghz",,,,,, ,"""-O2 no-mmx no-sse”","""-O2”","% Diff Between -O2s","“-O3 no-mmx no-sse”","“-O3”","% Diff Between -O3s" "File = mozilla","MB/s","MB/s",,"MB/s","MB/s", "LZ4",1588.5,1588.5,100.00%,1586.7,159.5,10.05% "ZFS LZJB",415.3,410.5,98.84%,415.8,410.7,98.77% "BSD LZJB",408,418.2,102.50%,382.8,391.9,102.38% "HAX LZJB",649.6,650.7,100.17%,653.5,656.7,100.49% ,,,,,, "% Diffs",,,,,, "HAX LZJB → LZ4",40.89%,40.96%,,41.19%,411.72%, "HAX LZJB → ZFS LZJB",156.42%,158.51%,,157.17%,159.90%, "HAX LZJB → BSD LZJB",159.22%,155.60%,,170.72%,167.57%, "BSD_LZJB → ZFS LZJB",98.24%,101.88%,,92.06%,95.42%, changes are in my github. Based on this all main future testing will be done with "fullbenchK" to try and get as close to kernel performance measurements as I can. If it turns out there are more "de-tunings" that should be in the "K" versions i can incorporate those as well, as required. Steven
_______________________________________________ developer mailing list [email protected] http://lists.open-zfs.org/mailman/listinfo/developer
