Re: Slow-down every 30 secs

From: Arjan de Vet <[email protected]>
Date: Tue, 24 Feb 1998 18:50:36 +0100 (CET)

In article <199802241117.MAA19129@cvtstu.cvt.stuba.sk> you write:

>We're running squid 1.1.20 on FreeBSD-2.2.5-Stable, with 256 MB RAM
>and 12 GB (3 x 4 GB SCSI-UW) of disk space dedicated to the cache.
>
>Unfortunately, we're experiencing the serious slow-down of all
>connections approximately every 30 secs. - see the attached netstat -w2

I've seen this too on at least FreeBSD and BSD/OS. Both operating systems do
a sync every 30 seconds and during that time all other disk I/O seems to
block completely while the disks are 100% busy with writes. This not only
happens with Squid, also with other applications.

On FreeBSD you can set the update interval to a lower value (as explained in
another reply). On BSD/OS I've experimented with a 'while true; do sync;
sleep 10; done' program. Maybe some more intelligent way of syncing is
needed which does the pending writes not all at once.

>output. When we try to connect to the squid in the slow-down interval,
>it accepts the connection at port 3128, however it does not return
>any data (we don't receive even an error message "Invalid request").
>After approx. 4 seconds of slow-down, everything works OK for next 26 secs.
>During the slow-down interval, the amount of memory dedicated to network
>significantly increases (from approx. 40 % to 70 % - see below the output
>of netstat -m)
>
>This looks like squid operation is completely suspended every 30 sec.
>for some reason - which might be e.g. kernel's syncing.

That's it. You can also run systat -vmstat 1 and then watch the disk
transfers on your cache disk(s).

Arjan
Received on Tue Feb 24 1998 - 10:18:04 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:38:58 MST