It sounds like the system is limited by bytes/sec not by ops/sec since
the bottleneck seems to be disk bandwidth. It might be helpful to
compute bytes written per sec at the API level, and then estimate how
often each "API byte" is written to disk and read from disk due to
compactions. That would support some quick back of the envelope
calculations to do a performance smell test...
Thorsten
Jonathan Ellis wrote:
Thanks for looking into this. Doesn't seem like there's much
low-hanging fruit to make compaction faster but I'll keep that in the
back of my mind.
-Jonathan