> -----Original Message----- > From: john allspaw [mailto:jallspaw@xxxxxxxxx] > Sent: Monday, September 12, 2005 7:58 AM > To: Henrik Nordstrom; Squid Users > Subject: Re: disk partition locations ? > > > thanks for the replies, guys. > > yes, we're using squid in a farm with very high performance > requirement, as it's doing http acceleration for our origin servers. > we're replacing 2disk SATA with 6disk SCSI, and while I > gather there are diminishing returns with adding over 3 or 4 > spindles, having a decent cache size is probably second in > importance. > > we'll be going from 2 cache_dirs of 10Gb each to 6 > cache_dirs (one on each disk) with 5Gb each. > our cache_mem size is 2048mb, which puts us just over the > 10mb/1Gb mem-to-disk suggestions, but the boxes have 4Gb of As I understand it, squid only populates cache_mem with newly requested objects (not objects from disk), so if you restart Squid, that 2GB of RAM is going to lay fallow. Using a smaller cache_mem allows the OS to use the rest of the memory to cache disk hits. The 10mb/1GB mem-to-disk suggestion is how much ADDITIONAL memory squid uses (for overhead, and the like): "As a rule of thumb on Squid uses approximately 10 MB of RAM per GB of the total of all cache_dirs (more on 64 bit servers such as Alpha), plus your cache_mem setting and about an additional 10-20MB. It is recommended to have at least twice this amount of physical RAM available on your Squid server." > RAM in them, and we can add more if need be. we're doing > roughly 3000 req/sec across an 8 machine farm, and these SATA > drives are getting awfully hot, with over an 80% hit rate > (40% mem, 40% disk). the objects never change once they are > in cache except for small cases, in which case we make an > explicit PURGE. > > we see upwards of 80-100% disk utilization at peak, so > getting the faster/more disks (ext2 with noatime, btw) seemed > like a good idea. :) > ok, from what you're both saying, I might not have to worry > about where to put the 5Gb partitions on these drives, but > basically, it can't hurt. I've been going on the assumption > that the most important performance gains will come from 1) > the xtra spindles, and 2) the better seek times on the 15K > SCSI (versus 7200k SATA) drives. > > thanks, > john > > > Chris