Re: --enable-async-io code status

From: Duane Wessels <wessels@dont-contact.us>
Date: Wed, 06 Jan 1999 13:01:34 -0700

"Jeff Madison" writes:

>Gentlemen,
>
>I noticed this message this morning and thought a problem that I am having
>may be of interest to you. Below is a message that I posted to the list
>last night and describes a problem that many who use -enable-async-io are
>encountering. I run Solaris 2.6_x86 with all patches. See the message
>below:
>
>----Begin----
>I have been working with squid for about a year now and it has been great
>fun. I have come across an issue that is causing us some serious headaches
>at this time. We run 2 what I would call large caches, the systems are Quad
>XEON 400, 1GB or RAM, and 54GB Hardware Level RAID Arrays. The problem I am
>having is that for speed and load reasons I have compiled
>with -enable-async-io. With this option enable Squid appears to "forget" or
>loose track of what and how many items it has in it's cache directory. I
>saw a message some time back that indicated that if you deleted the swap.*
>files in the cache directory that Squid would be forced to rebuild the
>swap-state file when it was restarted. This proved to be true. I would
>start Squid and it would say it only had approx 60000 objects in the cache
>directory, hard to believe when it's using 40 GB of drive space. After
>stopping Squid, removing swap.* files from cache directory and restating
>Squid it would report that it had over 2 million objects after is was done
>rebuilding. The problem is that with this much data squid takes hours to
>rebuild and with the load we place it under it chokes and dies before it
>ever finishes rebuilding forcing a restart and restarting the rebuild. We
>are running Squid 2.0.Patch2. My question is has this bug been fixed and in
>what version or does anyone have any suggestions.
>----End----
>
>Any comments would be appreciated.

I can't address anything related to async-io. Much of that code is
still mysterious to me. :-(

Someone recently suggested an option to disable storing objects to disk
during the rebuild phase. This would likely help your situation, and
would be easy to implement.

Also, you might have a look at the 'max_open_disk_fds' option.
If squid has this many disk files open, it stops saving objects to
disk. This might help your cache make it through the really high
load times with better response times.

Duane W.
Received on Wed Jan 06 1999 - 12:51:13 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:43:55 MST