guoqingzhu wrote: > Hi, can anyone help tell me about this? > I want to test my raw disk IO performance using 'dd' or other tools like > vdbench. But I don't know which device file should be correct, or both OK. > What about the difference between these two file when I do dd to them? > In my opinion I think when using /dev/rdsk/xxx, read/write requests are > directly sent to device driver. Otherwise when using /dev/dsk/xxx, read/write > requests will be processed first by high level file system interface(such as > vfs) then sent to device driver. So the /dev/rdsk/xxx is more suitable to > test raw disk IO performance. Is this thought correct? > Thanks > Guoqing >
Your conclusion is correct: you should use "character special" (aka "raw") devices (e.g. /dev/rdsk/*) to benchmark raw disk I/O performance. The term "character special" is something of a historical anomaly when talking about raw disk devices, because these devices require I/O to be block aligned. The "block special" devices (e.g. /dev/dsk/*) use mapped I/O interfaces as used by filesystems such as UFS (but not ZFS). As such, they use the system's page cache, so reads may be cached and writes will be asynchronous (unless synchronous writes are explicitly requested). However, both "character special" and "block special" devices are implemented using a virtual filesystem called "specfs" (in Solaris, the term "filesystem" is used to cover many things other than the likes of UFS, TMPFS, PCFS and ZFS). On the issue of buffering, I am quite often invited to comment on benchmarks and their results. I was recently shown sum stats from a test called "iozone". The results showed a sustained read performance in excess of 2GB/sec. However, when I asked about the storage configuration I was told that the storage was connected via two 4Gb/sec HBAs. The customer had not noticed that two 4Gb/sec HBAs are incapable of delivering 2GB/sec! I am not a great fan of dd for serious I/O benchmarking (although it does have its uses). You might like to check out Filebench, available via the OpenSolaris website. Hope this helps, Phil p.s. Merry Christmas from the UK! _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
