barel@imag.UUCP (Max BAREL) (07/28/87)
Since my last posting generate some useful answers, let's try again. When we get our ICM machine, we were quit happy but disapointed by speed of disk I/O. I wrote a simple benchmark, doing random BLOCK read on /dev/dsk/0s7, in order to minimise the buffer cache mechanism influence. This test ends up with a throughput of 10 kbyte/s. Our hard was: -emulex controller -CDC wren II 85 Mb disk The same test on an ATT 3b2/400 give 30Kb/s. I recently change the emulex cont. for an Adaptec RLL 2/7. This speed up the speed to 15Kb/s. Here is the question: What is slow in the chain? The ICM SCSI channel and the controler are specified for much more. We try other disk (hitachi,...), whithout succes. All have raisonable seek time. The driver could be the problem ? I hear my modest bench fly to NSC, which get much better result. It seem it is with a new release of the driver. I should welcome some light around this, since I can't believe the driver was so bad. Thank's. -- ------------------------- Max Barel barel@imag.UUCP
joer@nscpdc.NSC.COM (Joe Rawlings) (08/04/87)
Edited for brevity sake: In article <1879@imag.UUCP> barel@imag.UUCP (Max BAREL) writes: >When we get our ICM machine, we were quite happy but disapointed by speed of >disk I/O. >I wrote a simple benchmark, doing random BLOCK read on /dev/dsk/0s7, in order >to minimise the buffer cache mechanism influence. This test ends up with a >throughput of 10 kbyte/s. >Our disk environment was: > -emulex controller > -CDC wren II 85 Mb disk >The same test on an ATT 3b2/400 give 30Kb/s. >I recently change the emulex cont. for an Adaptec RLL 2/7. This speed up >the speed to 15Kb/s. >Here is the question: What is slow in the chain? >The ICM SCSI channel and the controler are specified for much more. >I should welcome some light around this, since I can't believe the driver >was so bad. >Thank's. > Disk throughput is dependent upon many things, however..... to find out why your figures are so low, we need to know how you are testing the throughput. We have tested the throughput and the following are peak and average values for both raw and block I/O in bytes per second using a Quantum Q160 disk drive. read write overall RAW I/O PEAK 489,411 409,600 AVERAGE 374,491 409,600 391,259 BLOCKED I/O PEAK 56,320 297,890 AVERAGE 25,458 105,931 41,051 Please recall that the possible throughput of a bus (SCSI in this case) has very little to do with the actual throughput, as OS/device command passing, syncronization, OS overhead, mechanical movements etc. must occur in addition to the actual data transfer. -- /****************************************************************************** * * * Joe Rawlings nsc!nscpdc!joer * * ICM Product Support tektronix!reed!nscpdc!joer * * National Semiconductor Corp. 1-800-222-2433 (outside Oregon) * * Portland Development Center (503)-629-4414 (Oregon/World) * * * ******************************************************************************/