Closed mkeeter closed 3 weeks ago
Here's main
running a 30-seconds crutest rand-write
with 4 KiB blocks. Notice that it takes an additional 30 seconds to actually finish the work; if someone sent a read
at that point, it would be delayed by up to 30 seconds.
Here's this branch, which finishes the jobs at a reasonable speed:
Closing in favor of a larger stack of small changes
Testing with
crutest bufferbloat
reveals that we can get multi-second delays if we send a single read after filling up the write queues.This PR turns up backpressure gain to make steady-state queue lengths less egregiously long.
I've done some casual testing, and now see turnaround times in the 10-100s of milliseconds. @faithanalog it would be great to get an
iodriver
run on this PR as well.