Re: [SQU] Memory usage breakdown (and MEM_REQUEST_T/request_t memory)

Are your web site by any chance running apples that pull content from
the server using periodic POST requests with a quite short interval?

Why I ask: It looks like host-b has a unnatural number of request
structures, and there is a known bug in Squid-2.3 and earlier in the way
POST requests are handler, causing them to lock up memory until the
client connection has been closed.

Workaround: Disable persistent connections

Fix: http://squid.sourceforge.net/hno/patch-2.3.html or
http://squid.sourceforge.net/hno/patch-2.2.html

--
Henrik Nordstrom
Squid hacker
James R Grinter wrote:
> 
> I've been looking at two Squid systems which are identical in all
> respects (exactly the same Solaris OS build, patches) including Squid
> 2.2 STABLE 4, and configuration installed from the same package (as
> built by myself.)
> 
> They're configured as accelerators, have been running for the same
> amount of time, and have serviced approximately the same number of
> requests. They are both utilising redirector processes. But their
> memory usage is radically different.
> 
> Here are some key values, taken from the process tables and from the
> 'mem' cache info.
> 
>           RSS   Storage   Store  Storage  StoreEntry SM
>                   swap   entries   Mem     memory    buf
> host-a   255MB 1349921kB  105024 124696kB  5747kB   124940kB
> host-b  1098MB 1166827kB  131280 614360kB  7183kB   796248kB
> 
>         Hotcache Store Entries     on-disk  Mem     request_t
>          items   with Mem Objects  objects  obj
> host-a   23747      23753           80657  2694kB       189kB
> host-b    1903      73218           69759  8298kB     80146kB
> 
>          short    medium  long
>         strings  strings strings
> host-a   7265kB    326kB   17kB
> host-b  42816kB  26664kB 6077kB
> 
> Both hosts have 2GB of RAM (cache_mem is set to 600MB), and they're
> running well within the memory of the machine. Both have a max
> cache_dir size of 2560MB, and are running with the default 13kB
> average object size estimate (stats report approx 16.75kB on each.)
> 
> It's the big descrepancy between the numbers of "Store Entries with
> Memory Objects" and the resultant memory consumption (host-b being
> over 3 times the size of host-a, tying up approx 600MB/780MB of RAM in
> Storage Memory/Storage Memory Buffers) that puzzles me.
> 
> Why are so many more objects stored in memory on host-b (and what is
> the difference between that and the Hostcache items)? Why might there
> be such a large amount of memory allocated to request_t on host-b
> compared to host-a? Both had an approximately equal number of current
> sessions at the time when I took these figures.
> 
> Are there any other numbers I should look at? (I have the full snapshot
> output of the 'info' and 'mem' cache_info pages for both systems.)
> 
> James.
> 
> --
> To unsubscribe, see http://www.squid-cache.org/mailing-lists.html
--
To unsubscribe, see http://www.squid-cache.org/mailing-lists.html