(This also may be an issue in Squid-3, as the mem_node class stores the data
inline.)
I did a little poking with the FreeBSD allocator - it allocates non-tiny
objects on page boundaries with a size rounded up to the next page.
[adrian@sarah ~]$ ./test1 test1 131072
allocating 12, then 4096 byte structures 131072 times..
RSS: 536868
[adrian@sarah ~]$ ./test1 test2 131072
allocating 4108 byte structure 131072 times..
RSS: 1059244
The source is available here:
http://www.creative.net.au/diffs/mem-size.c
I believe newer allocators do this sort of thing. Could people on other platforms
(Linux, Solaris, Windows? :) give the above a whirl and let me know what the
output is?
I'm going to revisit my (much) earlier work which folded the 4k buffer into the
mem_node and look at unwinding it. Its easy to do everywhere -except- the write
path and thats all very dirty.
That -should- drop the memory usage for large memory caches a bit.
It'll increase the allocator usage a bit but then there are other
much larger abuses of the allocator out there.
Comments?
Adrian
-- - Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support - - $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -Received on Tue Apr 22 2008 - 13:39:21 MDT
This archive was generated by hypermail 2.2.0 : Wed Apr 30 2008 - 12:00:07 MDT