On Wed, Feb 13, 2008 at 4:07 PM, Erez Zilber <[EMAIL PROTECTED]> wrote: > > I'm not familiar enough with xdd, so I don't know how to tune it. Can > you try the following commands (assuming that /dev/sg4 is your sg device)? > > READ: sgp_dd if=/dev/sg4 of=/dev/null bs=512 bpt=1024 count=20480000 > thr=8 time=1 > WRITE: sgp_dd of=/dev/sg4 if=/dev/zero bs=512 bpt=1024 count=20480000 > thr=8 time=1
You have been running tests with indirect (=buffered) I/O, so your tests say more about how well Linux can buffer data than about iSER and STGT. I have ran the following tests (target: 2 GB RAM disk, SDR 4x network, iSER, ib_write_bw: 933 MB/s): $ sgp_dd dio=1 if=/dev/sde of=/dev/null bs=512 bpt=1024 count=4194304 thr=8 time=1 time to transfer data was 3.896344 secs, 551.15 MB/sec 4194304+0 records in 4194304+0 records out $ sgp_dd dio=1 of=/dev/sde if=/dev/zero bs=512 bpt=1024 count=4194304 thr=8 time=1 time to transfer data was 2.858662 secs, 751.22 MB/sec 4194304+0 records in 4194304+0 records out Note: my opinion is that the parameters passed in the above tests are highly unrealistic with regard to predicting filesystem performance. The parameters passed to sgp_dd specify that all data is passed in units of 512 KB and with eight threads at a time. A typical filesystem communicates data over iSCSI in blocks of 4 KB to 32 KB. Bart Van Assche. _______________________________________________ Stgt-devel mailing list [email protected] https://lists.berlios.de/mailman/listinfo/stgt-devel
