Re: [UNIXS] Proy servers (fwd)

On Fri, 15 Oct 1999, Stephen R. van den Berg wrote:

> Duane Wessels wrote:
> >Here's the condensed version:
> > Squid 2.2.STABLE5 on Linux with async I/O
> > 50 req/sec
> > 55% hit ratio
> > 1.52 sec mean response time
>
> A tad bit high mean response times, I think.
> How does that compare with these real-life results?

The datacomm-1 workload has much higher delays for cache misses
than you (and I) observe in real-life. Your median miss delay
is about 470 milliseconds. The benchmark, on the other hand,
has 3181 milliseconds.

Here's Squid's MEAN/MEDIAN HIT/MISS response times from the
polygraph run (all times in milliseconds):

                  HITS MISS
        MEAN 169 3177
        MEDIAN 59 3181

>
> [ Note: The byte hit ratios are a bit off now, at peak usage, they're
> about 30%-35%.
> I know median-service-time is not the same as mean-service-time,
> but the minimum service time for a hit, is around 0.053s,
> comparing this with the dcomm-1 results, I see Cobalt specify
> a mere 1.98s; what did your test come up with?
> And why is it so much higher?

Squid's is not higher than Cobalt's. If you want to compare
the numbers in the "executive summary" table of section 2:

                      MEAN Response Time
                          (seconds)
         Throughput HIT ALL MISS Hit Ratio

            50 0.169 1.52 3.18 54.97

Duane W.
Received on Fri Oct 15 1999 - 12:15:46 MDT