Do some system-level profiling runs (oprofile under Linux, dtrace under Solaris) during the fill phase, the disk intensive phase and the disk overload phase. Are you graphing statistics? Can you graph stuff like CPU, swapping/paging, disk IO? Whats cache_mem set to? Adrian On Thu, Nov 08, 2007, Dave Raven wrote: > Hi all, > I'm busy testing a squid box with 8xSATA drives, 4gig of DDRII > memory and 2x 2.6gig dual core processors. I'm using the basic datacomm test > from polygraph. I've configured 6 of the drives to use COSS, and the other > two diskd (I've also done basic ufs tests). During all of the tests below > I'm using 96mb for cache memory... > > If I run 600RPS the unit handles it fine, for about 23 minutes - at that > stage network connections start rising very quickly (and eventually running > out) and responses on the polygraph client slow down until it dies. If I use > 1200 RPS it happens after about 10 minutes. > > Then if I use a single IDE drive (just using ufs), at 600RPS it handles fine > for about 14 minutes at which stage it dies. > > My main question is what is it that's causing the connections to rise. To me > it's that the responses are taking longer to fulfill - the reason for this I > assumed would be the disk drive. But how is the IDE drive going for so long? > 96mb of memory is filling up way faster than that, and I can see it > accessing the drive. The transfers per second and transfer speeds on the > drives don't change when it begins to fail, and neither do any real squid > stats... > > I've also tested just having a 500mb cache on one IDE drive, filling it > first and then doing this - it also lasts just as long (having to delete > files as well etc)... > > Any idea what's happening at that stage? > > Thanks for the help > Dave -- - Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support - - $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -
