Closed romange closed 9 months ago
Attention: 5 lines
in your changes are missing coverage. Please review.
Comparison is base (
9e64c47
) 79.00% compared to head (188609d
) 78.50%.
Files | Patch % | Lines |
---|---|---|
util/fibers/uring_socket.cc | 58.33% | 5 Missing :warning: |
:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
For same 3 configurations with GETs on empty database (testing the networking + 1 single hop, noop command):
memtier_benchmark -s 172.31.36.182 --ratio=0:1 -t 16 -c 18 -n 60000 --hide-histogram --distinct-client-seed
============================================================================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency p99 Latency p99.9 Latency KB/sec
----------------------------------------------------------------------------------------------------------------------------
Gets 740704.48 0.00 740704.48 0.38514 0.38300 0.55100 0.66300 28853.47
============================================================================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency p99 Latency p99.9 Latency KB/sec
----------------------------------------------------------------------------------------------------------------------------
Gets 716695.29 0.00 716695.29 0.40247 0.40700 0.59100 0.75100 27918.21
============================================================================================================================
Type Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency p99 Latency p99.9 Latency KB/sec
----------------------------------------------------------------------------------------------------------------------------
Gets 689701.00 0.00 689701.00 0.42315 0.43100 0.61500 0.74300 26866.67
wait_for_cqe
as small as possible to increase the chances that another thread will notify the proactor in time. That should decrease the amount of false positives where proactor blocks and then immediately wakes up do to notifications via kernel.Performance tests: dragonfly is running on a 8-cpu server (aarch64), memtier client with 16 cpus:
memtier_benchmark -s 172.31.36.182 --command=PING -t 16 -c 18 -n 60000 --hide-histogram --distinct-client-seed
PING because we test pure networking stuff.Dragonfly v1.13
Dragonfly with helio changes upto commit 6b3c2c5d249ccb0f64da80d168bc35c574724fb4 (https://github.com/romange/helio/pull/193)
Dragonfly with helio changes including direct-fds and uring improvements