Am 19.07.2017 um 23:16 schrieb Sitsofe Wheeler: > Hi, > > On 19 July 2017 at 18:22, Danny Al-Gaaf <danny.al-gaaf@xxxxxxxxx> wrote: >> >> I'm running into an issue with performance measurements on an iSCSI >> volume. The volume has set a QoS value of max. 30.000 iops. >> >> I run tests with the libaio engine against the volume with the following >> base parameter: >> >> fio --filename=/dev/sde --ioengine=libaio --direct=1 --buffered=0 >> --atomic=1 --bs=4k --invalidate=0 >> --readwrite=[read,write,randwrite,randread] --iodepth=128 --time_based >> --runtime=1200 --name=ft --write_iops_log=w_iops_log >> --write_bw_log=w_bw_log --write_lat_log=w_lat_log --output=fio_log >> >> With randread/randwrite everything is as expected. The volume reaches >> the QoS max value and not more. But with write/read/readwrite I can get >> over the max QoS IOPs: ~41k iops. >> >> After some investigation with wireshark it seems with seq r/w packages >> are aggregated (e.g. 8k TCP packages instead of 4k), which they are not >> with randwrite. >> >> Is there an issue in libaio or in fio? Or is it an expected behavior >> from the kernel? > > Is fio's output (see "And finally, the disk statistics are printed" in > http://fio.readthedocs.io/en/latest/fio_doc.html#interpreting-the-output > ) saying that merges took place? No, it says merge=0/248086. Thats why I was puzzled with the situation. It seems there is a small bug in fio. I will take a look at it. > If so if you tell the kernel to > disable all merging (see nomerges in > https://www.kernel.org/doc/Documentation/block/queue-sysfs.txt which > talks about how it is possible to disable merging by writing 2 to the > block device's nomerges parameter) does that change things? If so then > I'd say you are seeing expected kernel behaviour. By setting nomerges to '2' the behavior is changed and the performance fits the expectations. Thanks for the hint. Danny -- 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