17% throughput regression in insert workload

XMLWordPrintableJSON

    • Type: Bug
    • Resolution: Incomplete
    • Priority: Major - P3
    • None
    • Affects Version/s: 3.6.0
    • Component/s: Storage, Write Ops
    • None
    • Replication
    • Storage 2017-11-13, Storage 2017-12-04, Storage 2018-01-29, Repl 2018-02-26, Repl 2018-03-12, Repl 2018-03-26
    • 0
    • None
    • 3
    • None
    • None
    • None
    • None
    • None
    • None
    • None

      Simple insert workload:

          res = benchRun({
        ops: [
            {op: "insert", doc: {},  ns: "test.c"},
        ],
        seconds: 45,
        parallel: 15
    })
    printjson(res)

      standalone shows ~11% throughput reduction

      3.4.9:     "totalOps/s" : 312257.31243007537,
3.6.0-rc0  "totalOps/s" : 277214.2306740861,

      1-node replica set shows ~50% throughput reduction

      3.4.9:     "totalOps/s" : 181808.21357151057,
3.6.0-rc0  "totalOps/s" : 91202.31599375796,

      Metrics for the 1-node replica set case suggest issue may be related to large number of write-lock calls, about 3 per insert:

        1. ensure-session.patch
          2 kB
        2. pmp.out.3513
          17 kB
        3. pmp.out.359
          8 kB
        4. pmp.patch.sesscache
          12 kB
        5. run2.png
          run2.png
          127 kB

            Assignee:
            [DO NOT USE] Backlog - Replication Team
            Reporter:
            Bruce Lucas (Inactive)
            Votes:
            3 Vote for this issue
            Watchers:
            45 Start watching this issue

              Created:
              Updated:
              Resolved: