Re: Proxy of large files

From: David Luyer <luyer@dont-contact.us>
Date: Wed, 22 Jan 1997 11:47:03 +0800 (WST)

On Tue, 21 Jan 1997, Jonathan Larmour wrote:
>At 08:19 21/01/97 -0800, Duane Wessels wrote:
>>Does anyone have good or bad things to say about the 'NOVM' version
>>of squid-1.1.4? Does anyone think that switching over to the disk-only
>>approach is the wrong thing to do?
>
>I would be hesitant until I've heard glowing reports from big sites using
>the NOVM version, as the only negative point really is performance. Until
>someone can say for sure it works well under load, then I'd delay any
>changeover. All IMHO BTW.

Under modern OS, Linux especially, I would hope for a performance
improvement with the NOVM version. The buffering/caching in the OS is
probably better and more efficient than keeping two or more sets of
indexes for copies the same data, and possibly the same data in the
cache/buffer and squid memory cache. It is for a similar reason that many
people using Linux for news servers use NNRP_DBZINCORE_DELAY -1 and
INND_DBZINCORE 0; the OS keeps a copy of 'hot' files in RAM, so there's no
point in having your own copy in RAM and being swapped out, etc - this
just wastes RAM/swap.

Of course, this depends how you tune your buffer/cache/swapping
priorities, etc, and how good other operating systems are at buffering.
But for Linux at least, it seems a good idea, and instead of allocating
more memory to the cache, you just keep more RAM "free" (like on our news
server, where we can have 150M of RAM used in buffers/cache) and the OS
should choose the most used files on the system to cache (as long as it
is tuned to swap aggressively so it gets rid of all those idle processes
like getty).

David.
Received on Tue Jan 21 1997 - 19:54:55 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:34:08 MST