> Hi, > > I reviewed Squid filemap code and it's clear that in some cases > a large cache will have high CPU load. > filemap.c function file_map_create function starts with 2^13 elements > and expands element number only after the list is full. > So for example if cached objects are always slightly below 2^23 and bitmap > size is 2^23 it will take a lot of CPU to find next free bit. > > Itzcak Just an idea... But it looks possible at add a binary chop between the word-detection loop and the bit-detection loop to chop the word and seed the 'bit' variable. Thats a gain of n/2 tests immediately. /2c Amos > > On Mon, Oct 6, 2008 at 1:05 PM, Henrik Nordstrom > <henrik@xxxxxxxxxxxxxxxxxxx> wrote: >> On sön, 2008-10-05 at 16:38 +0200, Itzcak Pechtalt wrote: >>> When Squid reach several millions of objects per cache dir, it start >>> to be very CPU consumer, becuae every insertion and deletion of object >>> takes long time. >> >> Mine don't. >> >>> On my Squid 80-100GB had the CPU consumption effect. >> >> That's a fairly small cache. >> >> The biggest cache I have been running was in the 1.5TB range, split over >> a number of cache_dir, about 130GB each I think. >> >> But it is important you keep the number of objects per cache_dir well >> below 2^24. Preferably not more than 2^23. >> >> >> What I think is that you got bitten by something else than cache size.. >> >> Regards >> Henrik >> >
