Amos Shapira wrote:
> I experience the same problems with identical versions (1.1.18, Linux
> 2.0.32, libc6, but also with older versions of everything, including
> with libc5 (for those who still suspect libc6)). Curiously enough,
> even though I saw some traffic about this in this forum, the problem
> have not been addressed and seems to be ignored.
The developers are well avare about the problem. It is caused by memory
fragmentation causing a lot of free memory to be "useless". What we lack
is good ideas on how to avoid it in Squid, and time to implement the
changes needed.
Various malloc implementation do behave differently, and we are trying
to determine which one (of the free implementations) that handles
Squid's malloc pattern best, and which ones that doesn't.
Other things that cause this is how pending objects is handled in 1.1.X,
if one user requests one very big object (for example a server push that
never ends) then Squid allocates the same amount of memory. This memory
does cause the squid image to be large in size but keeping a small
active set, and a large free list.
The current recomendation is to
1. Add a lot of swap. In many cases most of this memory can be
swapped out without any noticeable performance penalty.
2. Monitor your system. If squid grows without stabilizing at less
than 2-3 times the expected size, or begins to page a lot (on a
system with enought memory) then you/we have a problem.
--- Henrik Nordstr�mReceived on Tue Jan 06 1998 - 17:56:29 MST
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:38:20 MST