> Recently I've spent a fair bit of time benchmarking a Squid system > whose COSS and AUFS storage (10GB total) + access logging are on a > RAID0 array of two consumer grade SATA disks. For various reasons, I'm > stuck with RAID0 for now, but I thought you might be interested to > hear that the box performs pretty well. I don't think anyone will be interested in RAID0, as Squid's simultaneous access of each cache_dir on different disks is loosely analogous to RAID0. RAID1 on the other hand is very interesting. > Some initial experiments suggest that removing RAID doesn't > particularly improve performance, but I intend to do a more thorough > set of benchmarks soon. Following on from my comment above, a single 20gig RAID0 cache_dir is probably not that much different to two 10gig cache_dirs on single disks. If using aufs then the RAID0 would only run as a single thread so that may adversely affect performance. I guess that RAID0 would offer a worse seek time from squid's perspectiv as each request from squid is serialised, but data transfer rate will be higher for a particular object. I imagine squid is more sensitive to seek than throughput. I'm just speculating on all this... Also, are you using the noatime mount option with reiserfs? Do you know what your "600 - 700 Req/Sec Polygraph polymix-4 benchmark" is in Mbps?
