I've noticed something odd when working on microbenchmarks
for the cache: I create and start loading channels in a loop,
then use the nsITimedChannel interface to get load-time for
each channel.

The time-series I get very often group measurements into
six consecutive similar values, then six higher but still
similar values, and so on. E.g. a time-series can look
like the following (numbers simplified for clarity):

1 2 1 1 2 1   4 6 5 4 4 6   9 9 8 7 9 8 ....

This happens when using a handler with httpd.js but also
when channels load from a separate python-server on localhost.
Note that both "network.http.max.connections" and
"network.http.max.connections-per-server" are set to 256 and
should not be the problem.

It does *not* occur second time I run the loop, loading all
resources from cache; values here are pretty consistent with
relative stddev of around 1%. I.e. it is related to loading
from the "net" (as opposed to loading from a cache).

The effect is also seen with all cache-devices disabled, i.e.
with no caching at all, hence the assumption is that this is
unrelated to the cache. (I'll verify this but need to decide
how first - it could be related to *writing* to cache.)

Any ideas what can cause this? Is there a limit in the OS
(Linux in my case) for concurrent local connections? I'm a
little puzzled over this and it pollutes results from the
benchmarks.
_______________________________________________
dev-tech-network mailing list
[email protected]
https://lists.mozilla.org/listinfo/dev-tech-network

Reply via email to