Re: apc improvements

Gonzalo Arana wrote: > Hi, > > I've tested apc in one of my apaches, and I've got this results: > > context switches (from vmstat) > w/out apc: 4k/sec > with apc: 200k/sec > > running processes (from 'vmstat 1') > w/out apc: 0-9 > with apc: 20-100 > > Cpu usage > with aps: 30% user, 60% system > w/out aps: 50% user, 10% system > > I suspect this figures may be the result of the fact that apc uses only > a rwlock (it does not implement nor use rw/ro locks). > > Therefore, internal structures are always locked in a mutex-style, even > in the case of a hit. > > In my opinion, apc needs a read-write/read-only locking strategy to > boost performance. With this implemented, in case of a hit to apc's > cache, the running process would not be rescheduled by the kernel. > Right now, if the lock is held by another process (even in case of a > hit to apc's cache), the running process would get scheduled. > > I would like to work on this. Does a feature like this have any chance > to get into apc's official source code? But there is a write on every hit for counters and the last accessed timestamp so I am not sure where a read-only lock would be applied. -Rasmus