The following is a rough comparison between COSS and AUFS on a pair of reasonably busy forward (transparent) Squid-2.6 caches running on Linux. Both proxies are using the TPROXY and WCCPv2 functionality to intercept requests and provide end-to-end transparency.
Thanks to Steven Wilton for the statistics and his ongoing efforts with COSS.
The COSS code was introduced near the beginning of Week 33. Note the substantial drop in IOWAIT. This is running on a PIII-500, 512MB RAM, 3x9GB SCSI HDD cache.
The next images show the drop in read/write and time spent in IO.
disk 1 - remained AUFS (not wiped)
disk 2 - blanked, turned into COSS
disk 3 - blanked, turned into COSS
Finally, the hit rate dipped slightly but has recovered over time, partly due to the cache being filled but also due to ongoing work to improve COSS.
hit ratio
load average
request rate
Similarly, this proxy also underwent the upgrade to COSS. Thsi proxy is a P4 Xeon 3GHz, 2Gb RAM, 5 x 36Gb SCSI HDD cache.
cpu
disk 1 - remained AUFS (not blanked)
disk 2 - remained AUFS (not blanked)
disk 3 - AUFS, turned into COSS
disk 4 - AUFS, turned into COSS
disk 5 - AUFS, turned into COSS
hit ratio
loadav
request rate
