On Mon, 14 Mar 2011 18:12:27 +0530, Saurabh Agarwal wrote:
> Thanks Amos. I will try doing those different sizes tests.
>
> Some more observations on my machine. If I don't transfer those 200
> HTTP Files for the first time in parallel but sequentially one by one
> using wget and after this if I use my other script to get these 200
> files in parallel from Squid then memory usage is allright. Squid
> memory usage remains under 100MB. I think for the first time transfer
> there is even more disk usage like save the files to disk and then
> read all of them parallel from disk. Also I think there should be
> lots
> of socket buffer space being used as well by Squid for each client
> and
> server socket.
>
> Regarding cache_dir usage what do you mean by "one cache_dir entry
> per spindle". I have only one disk and one device mapped partition
> with ext3 file system.
The config file you showed had 3 cache_dir on that 1 disk. This is bad.
Each cache_dir has N AIO threads (16, 32 or 64 by default IIRC) all
trying to read/write from random portions of the disk. Squid and AIO
scheduling does some optimization towards serialising access to the base
disk, but that does not work well when there are multiple independent
cache_dir state handlers.
Amos
Received on Mon Mar 14 2011 - 21:36:40 MDT
This archive was generated by hypermail 2.2.0 : Tue Mar 15 2011 - 12:00:01 MDT