diff options
| author | Jens Axboe <axboe@kernel.dk> | 2021-08-26 12:59:44 -0600 |
|---|---|---|
| committer | Jens Axboe <axboe@kernel.dk> | 2021-08-26 12:59:44 -0600 |
| commit | 7b05bf771084ff788243b78f51bc2c820730951c (patch) | |
| tree | 7220ad3f581eaa5fbad9ad0d86cd23dc2c5878c4 /block/blk-mq-rdma.c | |
| parent | b6d2b054e8baaee53fd2d4854c63cbf0f2c6262a (diff) | |
Revert "block/mq-deadline: Prioritize high-priority requests"
This reverts commit fb926032b3209300f9dc454a36b8299582ae545c.
Zhen reports that this commit slows down mq-deadline on a 128 thread
box, going from 258K IOPS to 170-180K. My testing shows that Optane
gen2 IOPS goes from 2.3M IOPS to 1.2M IOPS on a 64 thread box.
Looking in detail at the code, the main culprit here is needing to sum
percpu counters in the dispatch hot path, leading to very high CPU
utilization there. To make matters worse, the code currently needs to
sum 2 percpu counters, and it does so in the most naive way of iterating
possible CPUs _twice_.
Since we're close to release, revert this commit and we can re-do it
with regular per-priority counters instead for the 5.15 kernel.
Link: https://lore.kernel.org/linux-block/20210826144039.2143-1-thunder.leizhen@huawei.com/
Reported-by: Zhen Lei <thunder.leizhen@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Diffstat (limited to 'block/blk-mq-rdma.c')
0 files changed, 0 insertions, 0 deletions