> On Dec 3, 2018, at 3:20 PM, Noah Watkins <nwatkins@xxxxxxxxxx> wrote: > > On Tue, Nov 27, 2018 at 10:04 AM Yiming Zhang <yzhan298@xxxxxxxx> wrote: >> >> Hi All, >> >> I’m trying to understand the throttlers in Ceph, and I may need suggestions from the community. >> I plan to run some experiments against the throttler in the BlueStore to measure the latency performance. >> The parameters of interest include: >> >> bluestore_throttle_bytes >> bluestore_throttle_deferred_bytes >> bluestore_throttle_cost_per_io >> bluestore_throttle_cost_per_io_hdd >> bluestore_throttle_cost_per_io_ssd >> >> I’ll change these parameter and workload(size, type) to measure latency over time. The goal is to >> Propose a new throttler that can Keep as few operations in the BlueStore work queue as possible >> without increasing latency. >> >> Is this the right direction to go? >> Is there internal classifications for requests(onode, omap, collection obj, data blob) to be throttled in ObjectStore layer? > > I think you'll want to be careful how you construct the test, because > it seems like there is throttling / limits imposed at the bluestore > level, client level[, and probably in the OSD itself somewhere]. Are > you planning on testing BlueStore directly, or generating a workload > at the rados client level? I’m using rados bench to run the test, and with perf counter to observe the latency performance. The default throttler is in the BlueStore layer. Thanks, Yiming
