On 06/11/14 21:26, Nick Holland wrote: > On 06/11/14 15:55, Christian Weisgerber wrote: >> On 2014-06-11, Peter Fraser <[email protected]> wrote: > ... >>> Also for dd the block size has always been a puzzle. >> >> For accessing a raw device you want it to be a multiple of the >> sector size of the device (512 bytes for most disks) and there is >> usually no point in making it bigger than MAXPHYS (64k on OpenBSD), >> i.e., the maximal size of a single I/O transfer the kernel handles; >> larger reads or writes will be broken up into multiple transfers. > > I've heard this a number of times...and yet my testing on hardware I've > had in front of me (i.e., "your throughput may vary") has shown that > bs=1M does give substantially better throughput when zeroing disks than > 32k, and last time I did extensive testing in this, sizes larger than > 1MB give even better throughput, though the return gets very small after > around 1MB -- so I usually use 1MB so a "pkill -INFO dd" will give me an > indication of the progress in easy to read terms, which I find more > useful than a 1% reduction in time. > > I'm just reporting an observation, not explaining it. :) > > Nick. >
AAAANNNDDD... It was pointed out I missed the fact that my example (reading from /dev/zero) is quite different than reading from another disk. So...your results WILL vary... (I still like 1M block size for purpose of -INFO output ... but that wasn't the question) Nick.
