Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-16368

ZFS RPMs without vectorization

    XMLWordPrintable

Details

    • Improvement
    • Resolution: Unresolved
    • Minor
    • None
    • Lustre 2.15.1
    • RHEL8.6
    • 9223372036854775807

    Description

      The provided ZFS RPMs for RHEL86 Lustre servers are appearently built without vectorization support for raidz and fletcher4.

      The available algorithms are benchmarked on module load. The benchmarked implementations can be checked under

      cat /proc/spl/kstat/zfs/fletcher_4_bench
      implementation   native         byteswap
      scalar           5497759569     4845665644
      superscalar      5480937676     4858516672
      superscalar4     5488439650     4847575376
      fastest          scalar         superscalar
      
      cat /proc/spl/kstat/zfs/vdev_raidz_bench
      implementation   gen_p           gen_pq          gen_pqr         rec_p           rec_q           rec_r           rec_pq          rec_pr          rec_qr          rec_pqr
      original         254722784       200239743       97275027        1207670510      225639247       35724759        87910053        20776517        20811404        13297656
      scalar           1234755379      393704159       159539402       1211544896      385637841       266021529       195891453       153232648       90736440        81828326
      fastest          scalar          scalar          scalar          scalar          scalar          scalar          scalar          scalar          scalar          scalar
        

      The same hardware with official ZFS RPMs supports further optimizations, e.g.:

      cat /proc/spl/kstat/zfs/fletcher_4_bench
      implementation   native         byteswap
      scalar           6421841866     4605167866
      superscalar      9151122207     5740950514
      superscalar4     8273988438     6065318192
      sse2             10187627568    2365144114
      ssse3            10428775432    7506302223
      avx2             24481349937    19454027999
      avx512f          35018614756    10724551066
      avx512bw         36102325545    30780239769
      fastest          avx512bw       avx512bw

      A noticable performance improvement on high throughput scenarios could be observerd when changing to official ZFS RPMs with proper vectorization support.

      Attachments

        Activity

          People

            wc-triage WC Triage
            keller Patrick Keller
            Votes:
            0 Vote for this issue
            Watchers:
            1 Start watching this issue

            Dates

              Created:
              Updated: