Uploaded image for project: 'Couchbase Server'
  1. Couchbase Server
  2. MB-50250

PlasmaDGM tests regressed by 30%

    XMLWordPrintable

Details

    Description

      PlasmaDGM tests are regressed since 7.1.0-1887

      Test: 

      Secondary Scan Throughput (scanps) 1 bucket x 100M x 1KB, 2K KV ops/sec, 1000 indexes, 20% RR, Plasma, s=1 c=100

      7.1.0-1885 7.1.0-1887 7.1.0-1941
      205742.9 129,249.5 139555

      Latest run: http://perf.jenkins.couchbase.com/job/hemera/3536/console 

      Attachments

        Issue Links

          No reviews matched the request. Check your Options in the drop-down menu of this sections header.

          Activity

            Back to normal with 7.1.0-2238

            vikas.chaudhary Vikas Chaudhary added a comment - Back to normal with 7.1.0-2238

            This regression was seen in 2 tests:
            1) 1000idx scan throughput test with collections
            2) 1idx scan throughput test with bucket

            With fix in build couchbase-server-7.1.0-2238, I can see that the number for (1) is back to normal. But (2) still regresses. It still shows regression due to decompression overhead. Will need something like MB-50771 to help with it. The reason for the threshold changes not helping is that after init build pages are compressed and they are never swapped-out/swapped-in during scan phase. This is not the case with the 1000idx test where the working set seems to get swapped-in during scan phase and the threshold fix avoids compression here.

            akhil.mundroy Akhil Mundroy added a comment - This regression was seen in 2 tests: 1) 1000idx scan throughput test with collections 2) 1idx scan throughput test with bucket With fix in build couchbase-server-7.1.0-2238, I can see that the number for (1) is back to normal. But (2) still regresses. It still shows regression due to decompression overhead. Will need something like MB-50771 to help with it. The reason for the threshold changes not helping is that after init build pages are compressed and they are never swapped-out/swapped-in during scan phase. This is not the case with the 1000idx test where the working set seems to get swapped-in during scan phase and the threshold fix avoids compression here.

            Build couchbase-server-7.1.0-2238 contains plasma commit 98fa535 with commit message:
            MB-50250: Compress after swapin only if under memory pressure

            build-team Couchbase Build Team added a comment - Build couchbase-server-7.1.0-2238 contains plasma commit 98fa535 with commit message: MB-50250 : Compress after swapin only if under memory pressure

            Build couchbase-server-7.1.0-2238 contains indexing commit 5415ba9 with commit message:
            MB-50250: Add plasma CompressMemoryThreshold config

            build-team Couchbase Build Team added a comment - Build couchbase-server-7.1.0-2238 contains indexing commit 5415ba9 with commit message: MB-50250 : Add plasma CompressMemoryThreshold config

            Regression is due to decompression overhead. From build 7.1.0-1887 onwards, plasma will compress on swapin. If scans accesses only compressed pages, they will be slower due to decompression overhead. This can occur even when there is sufficient memory between usage and quota. To overcome this, compression on swapin can be performed only when there is memory pressure and with this patch, the performance for this test is back to normal. But, this can be bad for cases where the hot set does not change or gets swapped-out/swapped-in and the hot set does not fit in quota as in the case of the new compression tests. In these tests, regression is observed due to lower resident ratio from not compressing.

            akhil.mundroy Akhil Mundroy added a comment - Regression is due to decompression overhead. From build 7.1.0-1887 onwards, plasma will compress on swapin. If scans accesses only compressed pages, they will be slower due to decompression overhead. This can occur even when there is sufficient memory between usage and quota. To overcome this, compression on swapin can be performed only when there is memory pressure and with this patch , the performance for this test is back to normal. But, this can be bad for cases where the hot set does not change or gets swapped-out/swapped-in and the hot set does not fit in quota as in the case of the new compression tests. In these tests, regression is observed due to lower resident ratio from not compressing.

            People

              akhil.mundroy Akhil Mundroy
              vikas.chaudhary Vikas Chaudhary
              Votes:
              0 Vote for this issue
              Watchers:
              2 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved:

                Gerrit Reviews

                  There are no open Gerrit changes

                  PagerDuty