Thanks Jens. When you say neither number makes sense, are you referring to the sum of threads and the aggrb value, or the summed value and my suggestion of an average value? These threads are reading from a NAS device, so there's inherent variability expected. I suppose the ideal solution would be to read or write for a set amount of time, record the per-thread I/O done over that fixed time, then divide by the time. It seems like I could accomplish this using the "runtime" feature; the challenge (for me) with that is that I don't necessarily know the length of the test before the test is run, so it's difficult to choose an appropriate runtime. I didn't see anything in the documentation, but is there a way to stop all the threads in a group once one of the threads has completed all the requested I/O? On Thu, Oct 23, 2014 at 11:52 AM, Jens Axboe <axboe@xxxxxxxxx> wrote: > On 10/23/2014 11:04 AM, George Smith wrote: >> Uh oh, silence is never good :) Please let me know if I haven't >> included key information, missed something obvious, etc. >> >> Here's maybe a clearer example of what I'm talking about. The output >> file is from a read test using 5 threads: >> >> # grep 'read :' output.out >> read : io=51200MB, bw=116673KB/s, iops=113 , runt=449367msec >> read : io=51200MB, bw=189201KB/s, iops=184 , runt=277106msec >> read : io=51200MB, bw=143385KB/s, iops=140 , runt=365650msec >> read : io=51200MB, bw=114654KB/s, iops=111 , runt=457279msec >> read : io=51200MB, bw=183110KB/s, iops=178 , runt=286324msec >> >> # grep READ output.out >> READ: io=256000MB, aggrb=573269KB/s, minb=114653KB/s, >> maxb=189201KB/s, mint=277106msec, maxt=457279msec >> >> >> The sum of the threads is 747023, but aggrb is 573269. The bw= value >> in each thread line is the amount of I/O (from io=) divided by the >> time the I/O took. >> >> The aggrb= value is the total amount of I/O done (which is the sum of >> each thread's io= value), divided by maxt, which seems to be the >> maximum time seen during the run (which happens to be with my 4th >> thread). >> >> So it appears that this is the discrepancy. I'm not sure if it's >> correct to say the aggregate bandwidth is the total I/O divided by the >> max time that one of the threads in the group took to complete. Seems >> like taking the average time and dividing total I/O by that would be >> more correct. >> >> Am I missing the spirit of what the READ line is supposed to be conveying to me? > > I think you deduced it correctly. And yes, it is a bit misleading, in > that it differs from the sum of the reported bandwidths. Neither number > really makes sense, however. > > -- > Jens Axboe > -- To unsubscribe from this list: send the line "unsubscribe fio" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
