On 3 November 2016 at 01:58, Jaze Lee <jazeltq@xxxxxxxxx> wrote: > Is there some news for this bug ? > > 2016-11-02 11:50 GMT+08:00 Jaze Lee <jazeltq@xxxxxxxxx>: >> I found this is not related with rbd engine. If use libaio, and test >> raw device. It will also happen. >> the fio script is >> >> [global] >> ioengine=libaio >> filename=/dev/sda >> direct=1 >> group_reporting >> iodepth=32 >> thread >> >> [fio_iops_test] >> description=fio_iops_test >> runtime=600 >> bs=4K >> rw=randread >> write_lat_log=cpuiops >> rate_iops=2000 Can you make sure that you keep the fio mailing list CC'd on your mails rather than just replying to me directly? I'm just another fio user like you... I too can reproduce the problem with a slightly modified version of your job file: [fio_iops_test] ioengine=libaio direct=1 filename=/dev/nullb0 iodepth=2 bs=4K rw=read write_lat_log=/dev/null rate_iops=2000 Within a second or two of running it the following assert is printed: fio: stat.c:2004: get_cur_log: Assertion `iolog->pending->nr_samples < iolog->pending->max_samples' failed. Doing any of the following stops the assertion from triggering: setting the iodepth to 1, removing the write_lat_log parameter, removing the rate_iops parameter or adding the io_submit_mode=offload parameter. The problem seems to be due to the rate limiting somehow forcing iolog growth pressure but there is an inability to grow the iolog at the point we notice we need more room within stats. Normally the iolog growing can be deferred until next submission but I guess when you have to average over the samples in it (so as to work out the write latency for a given period) this is an issue because the logging areas aren't prepared for the change in size. You could hack the fio code to force a bigger default iolog size so it doesn't have to grow it later but a better way would be to use the io_submit_mode=offload parameter (https://github.com/axboe/fio/blob/fio-2.15/HOWTO#L888 ) which decouples the submission from the completion/logging. Perhaps fio could put up a warning suggesting its use if you use rate_iops and one of the parameters that force averaging together but there could be other ways of triggering the same issue... -- Sitsofe | http://sucs.org/~sits/ -- 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
