On Mon, 2010-07-26 at 14:34 -0400, Greg Smith wrote:
> Matthew Wakeling wrote:
> > Yeb also made the point - there are far too many points on that graph
> > to really tell what the average latency is. It'd be instructive to
> > have a few figures, like "only x% of requests took longer than y".
>
> Average latency is the inverse of TPS. So if the result is, say, 1200
> TPS, that means the average latency is 1 / (1200 transactions/second) =
> 0.83 milliseconds/transaction.
This is probably only true if you run all transactions sequentially in
one connection?
If you run 10 parallel threads and get 1200 sec, the average transaction
time (latency?) is probably closer to 8.3 ms ?
> The average TPS figure is normally on a
> more useful scale as far as being able to compare them in ways that make
> sense to people.
>
> pgbench-tools derives average, worst-case, and 90th percentile figures
> for latency from the logs. I have 37MB worth of graphs from a system
> showing how all this typically works for regular hard drives I've been
> given permission to publish; just need to find a place to host it at
> internally and I'll make the whole stack available to the world. So far
> Yeb's data is showing that a single SSD is competitive with a small
> array on average, but with better worst-case behavior than I'm used to
> seeing.
>
> --
> Greg Smith 2ndQuadrant US Baltimore, MD
> PostgreSQL Training, Services and Support
> greg@2ndQuadrant.com www.2ndQuadrant.us
>
>
--
Hannu Krosing http://www.2ndQuadrant.com
PostgreSQL Scalability and Availability
Services, Consulting and Training