- YCSB 30M documents, 10 fields, ~1kB/document, total ~30GB
- 50/50 read/update workload
- 40 GB cache, 128 GB memory, 32 CPUs
- slow SSD disk (~80-100 MB/s)
- no journal (to simplify the situation)
- per mongostat, cache is at 100% utilization, 80% dirty pretty much throughout the test.
During each checkpoint while a transaction id is pinned for the checkpoint throughput falls by an order of magnitude (but not to 0). This is accompanied by activity in the "checkpoint blocked eviction" counter. This is seen in B-C, F-G, J-K below:
Similar test with a larger cache (the default 64GB) also shows this issue.
Note: this is the same test as reported in
SERVER-18314; opening two separate tickets to track what may be separate issues.
- related to
SERVER-18314 Stall during fdatasync phase of checkpoints under WiredTiger and EXT4
SERVER-18677 Throughput drop during transaction pinned phase of checkpoints under WiredTiger (larger data set)