Add a background IOPS feature to the filesystem performance collector
that specifies separate read and write background IOPS to perform while
measuring latency. This allows for better assessment of whether etcd
will be stable when running alongside other workloads on the same
cluster.
Also add templating to the outcome message of the filesystem performance
analyzers to allow printing individual latency percentiles or the entire
table.
Remove the random IOPS benchmark since it was attempting to perform
unaligned direct I/O.
The random IOPS benchmark attempts to replicate the results of this
fio command:
fio --ioengine=psync --direct=1 --bs=4k --size=1Gi --readwrite=randread --serialize_overlap=1
Across three tests the fio command reported 1877 IOPS and the preflight
1822 IOPS with the same block and file size.
The included example found P99 latency of 2.6ms.
Fio reported P99 latency of 2.5ms with this command:
fio --rw=write --ioengine=sync --fdatasync=1 --directory=/var/lib/etcd
--size=220m --bs=2300
Andrew Reed