Due to the fact that queue_state_fiber called box.ctrl.wait_r* every millisecond, the idle queue consumed more than 10% of CPU.
The proposed solution makes calls to box.ctrl.wait_r* blocking until the read/write mode is changed.
Thus, CPU consumption is reduced to about ~1%.
Below is a performance comparison using a script t/benchmark/multi_consumer_work.lua whith params consumers-count = 650 and batch-size = 5000 on CPU AMD Ryzen 5 5600U.
For comparison, consider three commits.
The first column is commit (9ff105b) before implementing queue_state_fiber.
The second column is commit (159a43d) before implementation of the described proposal. That is, queue_state_fiber calls box.ctrl.wait_r every millisecond.
The third column (offered) is the current offer, the box.ctrl.wait_r blocking call.
The data is given as an arithmetic mean of 25 measurement points:
[time\commit]
9ff105b
159a43d
offered
fill queue
8 197 452
8 567 415
8 751 545
task confirm
49 817 371
52 203 080
53 020 243
Final performance comparison as a percentage:
% down
9ff105b->159a43d
159a43d->offered
9ff105b->offered
fill queue
4.513% down
2.149% down
6.759% down
task confirm
4.788% down
1.565% down
6.429% down
As a result, the proposed solution causes a decrease in performance for fill queue 2.149% and task confirm 1.565%.
Due to the fact that queue_state_fiber called box.ctrl.wait_r* every millisecond, the idle queue consumed more than 10% of CPU.
The proposed solution makes calls to box.ctrl.wait_r* blocking until the read/write mode is changed.
Thus, CPU consumption is reduced to about ~1%.Below is a performance comparison using a script
t/benchmark/multi_consumer_work.luawhith params consumers-count = 650 and batch-size = 5000 on CPU AMD Ryzen 5 5600U.For comparison, consider three commits. The first column is commit (9ff105b) before implementing queue_state_fiber. The second column is commit (159a43d) before implementation of the described proposal. That is, queue_state_fiber calls box.ctrl.wait_r every millisecond. The third column (offered) is the current offer, the box.ctrl.wait_r blocking call.
The data is given as an arithmetic mean of 25 measurement points:
Final performance comparison as a percentage:
As a result, the proposed solution causes a decrease in performance for fill queue 2.149% and task confirm 1.565%.Closes #183