Travel Factory S.r.l. wrote:
On a squid 2.7 Stable 6, 2 CPU 3.6 Ghz, 10 GB ram, 1 raid1 300gb disk,
partition /u02 is ext3 mounted with noatime, noatimedir, I have these two
cache_dirs - the setup for coss I took from a previous message :
cache_dir aufs /u02/squid 150000 256 256 min-size=4288
cache_dir COSS /dev/cciss/c0d0p5 38000 max-stripe-waste=32768 block-size=4096
maxfullbufs=10 max-size=524289
During normal use, I see that almost all objects are sent to the coss storage, and
only big ones ( > 550 KB ) are sent to the aufs storage. Actually it stores
flv, swf, jpg....
If I stop squid and restart it, during the time coss reads its stripes I get SO_FAIL error for objects less than 4288 bytes (and it is correct), and I have objects more than 4288 bytes long correctly stored to aufs...
From what I understand, these cache_dirs say:
- objects less than 4288 bytes ALLWAYS to coss
- objects more than 524289 ALLWAYS to aufs
- objects more than 4288 bytes AND less than 524289 split to aufs and coss....
But I'm probably wrong...
.. can you tell me what is wrong in my reasoning ? Because 524289 is probably
too big for 38 gb coss....
Thank you,
Francesco
PS: Mean Object Size: 30.65 KB
Without looking I'd guess that Adrian and the others who tuned COSS and
Squid-2.7 for high speed reads/writes did something to bias the storage
towards the most efficient cache_dire types available.
If I had to guess I'd say make the split size the same and put it around
128KB (about 4 times avg). IIRC COSS uses 1MB strips to 2 objects per
strip may or may not be good.
Amos
--
Please be using
Current Stable Squid 2.7.STABLE6 or 3.0.STABLE16
Current Beta Squid 3.1.0.9
