On Tue, 17 Aug 2010 22:43:33 -0430, Jose Ildefonso Camargo Tolosa <ildefonso.camargo@xxxxxxxxx> wrote: > Hi! > > In my own personal opinion: your hard drive alone is not enough to > handle that much traffic (110MBytes/s, ~1Gbps). See, most SATA hard > drives (7200rpm) gives around 50~70MB/s *sequential* read speed, your > cache reads are *not* sequential, so, it will be slower. In my > opinion, you need something like a 8 drives RAID10 array, and/or use > faster disks (10k), or maybe 15k SAS disks. > > Also, I would put a minimum object size for disk of 1M, and a maximum > object size of whatever you want (this depends on your network, but > usually ~150MB is enough to fit almost any upgrade download). And for > RAM, I would put a maximum object size of 1M, with no minimum. Thus, > keeping small files out of the disk cache. The COSS storage type he has setup already does this very efficiently with added disk-backing of the COSS chunks for cross-restart recovery of the cache. > > Also, other questions: How many clients/connections are you handling? > what are your server's specifications? and how is the system load over > time? (maybe zabbix or any other monitoring system will let you know > your system load over time). > > I hope this helps, > > Ildefonso Camargo > > On Tue, Aug 17, 2010 at 10:26 PM, Robert Pipca <robertpipca@xxxxxxxxx> > wrote: >> Hi. >> >> I'm using squid on a high speed network (with 110M of http traffic). >> >> I'm using 2.7.STABLE7 with these cache_dir: >> >> cache_dir aufs /cache 756842 60 100 Whats missing appears to be min-size=1048576 on the AUFS to push all the small objects into the better COSS directories. (NOTE: the value is COSS max-size+1) >> cache_dir coss /cache/coss1 65520 max-size=1048575 >> max-stripe-waste=32768 block-size=4096 membufs=15 >> cache_dir coss /cache/coss2 65520 max-size=1048575 >> max-stripe-waste=32768 block-size=4096 membufs=15 >> cache_dir coss /cache/coss3 65520 max-size=1048575 >> max-stripe-waste=32768 block-size=4096 membufs=15 >> >> Everything works fine most of the day, but on peak hours, I got these: >> >> 2010/08/17 20:06:59| squidaio_queue_request: WARNING - Disk I/O >> overloading >> 2010/08/17 20:06:59| squidaio_queue_request: Queue Length: >> current=981, high=1488, low=321, duration=170 >> >> After a while, I got a few of these, with "duration" increasing, until: >> >> 2010/08/17 20:23:09| squidaio_queue_request: WARNING - Disk I/O >> overloading >> 2010/08/17 20:23:09| squidaio_queue_request: Queue Length: >> current=558, high=2177, low=321, duration=531 >> >> The web browsing started to get very slow, which is when the support >> team took squid down. >> >> All cache_dir are on a single sata-2 7200RPM 1TB hard drive. >> >> Is there a way to know which cache_dir is the problem and what I can >> so this doesn't happen? AIO is the I/O method preferred by AUFS. That aufs dir is also listed first, which may affect the default choice. >> >> I tried using both 16 and 32 AIO threads, but didn't help. >> >> cache manager tells me that I have around 10 million objects: >> >> Average HTTP requests per minute since start: 18851.1 >> >> Storage Swap size: 693535688 KB >> Storage Mem size: 30872 KB >> Mean Object Size: 64.50 KB >> >> Internal Data Structures: >> 10752896 StoreEntries >> 49 StoreEntries with MemObjects >> 26 Hot Object Cache Items >> 10752847 on-disk objects *49* of 10 million objects are in-transit? that is very low. Though it could be a result of the queue overload. Amos