Uploaded image for project: 'WiredTiger'
  1. WiredTiger
  2. WT-2511

Recent WT builds have a large perf regression as thread count increases in read only workloads larger than cache

XMLWordPrintableJSON

    • Type: Icon: Improvement Improvement
    • Resolution: Duplicate
    • Priority: Icon: Major - P3 Major - P3
    • None
    • Affects Version/s: None
    • Component/s: None
    • None

      Comparison of the tip of develop (2.8.0) vs fb8739f shows a very significant performance regression (4x) when doing a read only workload.

      The following wtperf config can be used to reproduce:

      conn_config="eviction_trigger=81,session_max=20000,cache_size=3GB,eviction=(threads_max=4),checkpoint=(wait=60,log_size=2GB),log=(enabled=true,archive=true,path=journal,compressor=snappy),statistics=(fast,clear),statistics_log=(wait=1)"
table_config="type=file,leaf_page_max=32k,split_pct=90,memory_page_max=10M,leaf_value_max=64M"
compression="snappy"
icount=10000000
report_interval=5
table_count=2
run_time=120
populate_threads=1
threads=((count=100,reads=1))
sample_interval=5
value_sz=10000

      Checking stats the biggest thing I noticed that there seemed to be far more blocks (and bytes) being read in during the read phase of this test. I've uploaded the wtstats.py outputs to this ticket for develop (good.html) or fb8739f (bad.html).

        1. bad.html
          1.08 MB
        2. good.html
          990 kB

            Assignee:
            david.hows David Hows
            Reporter:
            david.hows David Hows
            Votes:
            0 Vote for this issue
            Watchers:
            8 Start watching this issue

              Created:
              Updated:
              Resolved: