On 3/6/18 4:16 PM, Jeff Furlong wrote: > Hi All, > Starting in fio 2.99, there is an odd problem when --ramp_time is used with --write_iops_log. The reported IOPs for the first ~200s or so are artificially inflated by fio, always monotonically increasing until about 200s, then resets to normal. With fio 2.99, disabling --ramp_time avoids the issue. And, with fio 2.21, using --ramp_time and --write_iops_log does not create the problem. Consider a job: > > # fio --name=test --ioengine=libaio --direct=1 --rw=randread --iodepth=1 --size=100% --bs=4k --filename=/dev/nvme1n1 --runtime=300s --ramp_time=10s --write_iops_log=test --log_avg_msec=1000 > > See the attached plot for fio induced IOPs during the first ~200s. Using a longer runtime makes identifying the issue easier if the DUT is consistently performing. > > One simple method to resolve this inflation is by reducing the LOG_MSEC_SLACK from 10 milliseconds to 1 millisecond, given that we now operate some time structures to the nanosecond level (so 10ms of slack seems excessive). Below I also fixup some data type issues to match existing types (avoids typecasts or potential compiler warnings): Thanks, seems reasonable enough to reduce the slack, 10ms is a lot. -- 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
