r583 - in trunk/varnish-cache: bin/varnishd include
Poul-Henning Kamp
phk at phk.freebsd.dk
Tue Aug 1 12:25:37 CEST 2006
In message <20060801093953.4C6CF1EC3EB at projects.linpro.no>, phk at projects.linpro
.no writes:
>SHMlog the start and difference between start and stop with ReqServTime
>tag.
I just ran a very quick live-test to see the distrbution of our
service time.
For both hits and misses we have a "hockey-stick" figure with
a small percentage of the transactions accounting for a large
part of the average.
The tables below show the numbers for the the entire dataset and
for the 99, 95, 90 and 75% fastest of the dataset.
Notice how rapidly the Max value falls off.
Since some of the hits may actually have been waiting on a busy
object while another session filled it from the backend, I think
it is fair to say that a "un-encumbered" hit takes less than
100 usec and typically less than 50usec to process.
Poul-Henning
===> hit
N Min Max Median Avg Stddev
100% 5196 24.865e-06 0.002863344 44.98e-06 54.471059e-06 46.992297e-06
99% 5144 24.865e-06 0.000150585 44.98e-06 52.087417e-06 16.672175e-06
95% 4936 24.865e-06 86.886e-06 44.701e-06 49.620993e-06 11.263975e-06
90% 4676 24.865e-06 77.387e-06 44.421e-06 47.891206e-06 8.7613014e-06
75% 3897 24.865e-06 51.406e-06 43.862e-06 44.476003e-06 2.5034017e-06
===> fetch
N Min Max Median Avg Stddev
100% 701 0.002255418 13.741159 0.010918089 0.21945862 0.87530516
99% 693 0.002255418 2.5146695 0.010593451 0.14724379 0.3537625
95% 665 0.002255418 1.0104925 0.009807282 0.086928877 0.18072753
90% 630 0.002255418 0.50737959 0.009042904 0.052588283 0.10405137
75% 525 0.002255418 0.089049749 0.006494698 0.013066963 0.015961684
--
Poul-Henning Kamp | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG | TCP/IP since RFC 956
FreeBSD committer | BSD since 4.3-tahoe
Never attribute to malice what can adequately be explained by incompetence.
More information about the varnish-commit
mailing list