On Thu, Jul 25, 2013 at 06:58:56AM +1200, Amos Jeffries wrote: > On 25/07/2013 1:05 a.m., Golden Shadow wrote: > >Hi there! > > > >My squid is installed on a server with 192 GB of RAM. I have the following > >directives in squid.conf: > > > >cache_mem 143360 MB > >maximum_object_size_in_memory 300 KB > >memory_replacement_policy heap GDSF > > > >memory_pools on > >memory_pools_limit 1024 MB > > > >ipcache_size 2048 > >ipcache_low 90 > >ipcache_high 95 > > > >fqdncache_size 2048 > > > > > > > >top reports that my squid process size is 20GB, which is far less than my > >RAM size, but nevertheless I still find some page faults (about 70 page > >faults over 2 hours). I'm wondering how could those page faults are > >occurring while squid process size is far less than my RAM size. How can I > >eliminate those time consuming page faults? > > Two things here. > > Why is the process size only 20GB? you have a 143GB memory cache as part > of that RAM consumption by Squid. Perhapse your traffics real caching > requirement is far smaller than you are allowing storage for. > > > What exactly is the page faulting comign from though ... Squid or the OS? > If it is Squid, why would the OS have swapped that piece of memory out > to VM in the first place? perhapse something else is needing a chunk of > memory larger than Squid leaves available? With plenty of memory and a very large process image, my guess is that the page fault is a TLB miss and not a swapin. TLB miss cannot be avoided if a process is larger than the TLB can address. Marcus > > >My second question, am I using correct values for the memory-related > >directives mentioned above? If no, I would really appreciate if you could > >suggest the correct values. > > Any values you want are "correct", so long as they fit within the > machines limits and do not lead to the system swapping. > > > >My last question is about read_ahead_gap, whose default value is only 16 > >KB. Would increasing this value to let's say 32 KB or 64 KB increase the > >performance since I have high RAM on the server? > > Perhapse. That is a buffer size more related to your network speed. Each > concurrent connection consumes up to that much RAM for buffers. If you > have clients that can drain 32KB or 64KB fast enough not to cause waves > or bursts in traffic it can be worthwhile raising it a bit. If you have > slow clients the reverse can be true. > > Amos
