On Tue, 20 May 2003, Bryn Reeves wrote:
> Hi all,
> Just a short query this time ;-) We have a cache which seems fairly
> un-stressed during an actual run (low cpu and disk utilization), but
> only achieves poor to average hit-ratios at the rates we'd expect.
>
> Would it be seem reasonable to think that the fill rate is set too high
> for this cache?
>
> We are currently using 75% and would expect to see good performance from
> this box at 75-100 reqs/sec minimum; however, although response times
> and errors are fine, we don't get near the target hit ratio until we
> bring the rate down to 25 reqs/sec.
Its possible that the fill rate is too high, but I'm skeptical.
During a normal polymix fill phase you should see 5% hit ratio. If
the fill rate is too high, such that the cache isn't caching
everything, you should see an average hit ratio lower than 5% during
the fill phase.
Another possibility is that the cache is too small. In a Polymix
test, the working set size is equivalent to about 4 hours of peak
traffic. So if your peak is 100 req/sec, the working set size is
about 1,440,000 objects (about 14 GB). Assuming the cache uses
LRU, it will need to be even larger to get the ideal hit ratio.
In our experience, it needs to be about 30% larger, so you probably
need at least 18GB to get the ideal DHR at 100 req/sec.
Duane W.
This archive was generated by hypermail 2b29 : Mon Feb 06 2006 - 12:00:26 MST