Re: [PATCH net-next v2 2/2] ptr_ring: make __ptr_ring_empty() checking more reliable

From: Yunsheng Lin
Date: Fri Jun 25 2021 - 05:20:17 EST


On 2021/6/25 14:39, Michael S. Tsirkin wrote:
> On Fri, Jun 25, 2021 at 11:18:56AM +0800, Yunsheng Lin wrote:
>> Currently r->queue[] is cleared after r->consumer_head is moved
>> forward, which makes the __ptr_ring_empty() checking called in
>> page_pool_refill_alloc_cache() unreliable if the checking is done
>> after the r->queue clearing and before the consumer_head moving
>> forward.
>>
>> Move the r->queue[] clearing after consumer_head moving forward
>> to make __ptr_ring_empty() checking more reliable.
>>
>> As a side effect of above change, a consumer_head checking is
>> avoided for the likely case, and it has noticeable performance
>> improvement when it is tested using the ptr_ring_test selftest
>> added in the previous patch.
>>
>> Using "taskset -c 1 ./ptr_ring_test -s 1000 -m 0 -N 100000000"
>> to test the case of single thread doing both the enqueuing and
>> dequeuing:
>>
>> arch unpatched patched delta
>> arm64 4648 ms 4464 ms +3.9%
>> X86 2562 ms 2401 ms +6.2%
>>
>> Using "taskset -c 1-2 ./ptr_ring_test -s 1000 -m 1 -N 100000000"
>> to test the case of one thread doing enqueuing and another thread
>> doing dequeuing concurrently, also known as single-producer/single-
>> consumer:
>>
>> arch unpatched patched delta
>> arm64 3624 ms + 3624 ms 3462 ms + 3462 ms +4.4%
>> x86 2758 ms + 2758 ms 2547 ms + 2547 ms +7.6%
>
> Nice but it's small - could be a fluke.
> How many tests did you run? What is the variance?
> Did you try pinning to different CPUs to observe numa effects?
> Please use perf or some other modern tool for this kind
> of benchmark. Thanks!

The result is quite stable, and retest using perf stat:

---------------unpatched ptr_ring.c begin----------------------------------

perf stat ./ptr_ring_test -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2385198 us

Performance counter stats for './ptr_ring_test -s 1000 -m 0 -N 100000000':

2385.49 msec task-clock # 1.000 CPUs utilized
26 context-switches # 0.011 K/sec
0 cpu-migrations # 0.000 K/sec
57 page-faults # 0.024 K/sec
6202023521 cycles # 2.600 GHz
17424187640 instructions # 2.81 insn per cycle
<not supported> branches
6506477 branch-misses

2.385785170 seconds time elapsed

2.384014000 seconds user
0.000000000 seconds sys


root@(none):~# perf stat ./ptr_ring_test -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2383385 us

Performance counter stats for './ptr_ring_test -s 1000 -m 0 -N 100000000':

2383.67 msec task-clock # 1.000 CPUs utilized
26 context-switches # 0.011 K/sec
0 cpu-migrations # 0.000 K/sec
57 page-faults # 0.024 K/sec
6197278066 cycles # 2.600 GHz
17424207772 instructions # 2.81 insn per cycle
<not supported> branches
6495766 branch-misses

2.383941170 seconds time elapsed

2.382215000 seconds user
0.000000000 seconds sys


root@(none):~# perf stat ./ptr_ring_test -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2390858 us

Performance counter stats for './ptr_ring_test -s 1000 -m 0 -N 100000000':

2391.16 msec task-clock # 1.000 CPUs utilized
25 context-switches # 0.010 K/sec
0 cpu-migrations # 0.000 K/sec
57 page-faults # 0.024 K/sec
6216704120 cycles # 2.600 GHz
17424243041 instructions # 2.80 insn per cycle
<not supported> branches
6483886 branch-misses

2.391420440 seconds time elapsed

2.389647000 seconds user
0.000000000 seconds sys


root@(none):~# perf stat ./ptr_ring_test -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2389810 us

Performance counter stats for './ptr_ring_test -s 1000 -m 0 -N 100000000':

2390.10 msec task-clock # 1.000 CPUs utilized
26 context-switches # 0.011 K/sec
0 cpu-migrations # 0.000 K/sec
58 page-faults # 0.024 K/sec
6213995715 cycles # 2.600 GHz
17424227499 instructions # 2.80 insn per cycle
<not supported> branches
6474069 branch-misses

2.390367070 seconds time elapsed

2.388644000 seconds user
0.000000000 seconds sys

---------------unpatched ptr_ring.c end----------------------------------



---------------patched ptr_ring.c begin----------------------------------
root@(none):~# perf stat ./ptr_ring_test_opt -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2198894 us

Performance counter stats for './ptr_ring_test_opt -s 1000 -m 0 -N 100000000':

2199.18 msec task-clock # 1.000 CPUs utilized
23 context-switches # 0.010 K/sec
0 cpu-migrations # 0.000 K/sec
56 page-faults # 0.025 K/sec
5717671859 cycles # 2.600 GHz
16124164124 instructions # 2.82 insn per cycle
<not supported> branches
6564829 branch-misses

2.199445990 seconds time elapsed

2.197859000 seconds user
0.000000000 seconds sys


root@(none):~# perf stat ./ptr_ring_test_opt -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2222337 us

Performance counter stats for './ptr_ring_test_opt -s 1000 -m 0 -N 100000000':

2222.63 msec task-clock # 1.000 CPUs utilized
23 context-switches # 0.010 K/sec
0 cpu-migrations # 0.000 K/sec
57 page-faults # 0.026 K/sec
5778632853 cycles # 2.600 GHz
16124210769 instructions # 2.79 insn per cycle
<not supported> branches
6603904 branch-misses

2.222901020 seconds time elapsed

2.221312000 seconds user
0.000000000 seconds sys


root@(none):~# perf stat ./ptr_ring_test_opt -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2251980 us

Performance counter stats for './ptr_ring_test_opt -s 1000 -m 0 -N 100000000':

2252.28 msec task-clock # 1.000 CPUs utilized
25 context-switches # 0.011 K/sec
0 cpu-migrations # 0.000 K/sec
57 page-faults # 0.025 K/sec
5855668335 cycles # 2.600 GHz
16124310588 instructions # 2.75 insn per cycle
<not supported> branches
6777279 branch-misses

2.252543340 seconds time elapsed

2.250897000 seconds user
0.000000000 seconds sys


root@(none):~#
root@(none):~# perf stat ./ptr_ring_test_opt -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2209415 us

Performance counter stats for './ptr_ring_test_opt -s 1000 -m 0 -N 100000000':

2209.70 msec task-clock # 1.000 CPUs utilized
24 context-switches # 0.011 K/sec
0 cpu-migrations # 0.000 K/sec
58 page-faults # 0.026 K/sec
5745003772 cycles # 2.600 GHz
16124198886 instructions # 2.81 insn per cycle
<not supported> branches
6508414 branch-misses

2.209973960 seconds time elapsed

2.208354000 seconds user
0.000000000 seconds sys


root@(none):~# perf stat ./ptr_ring_test_opt -s 1000 -m 0 -N 100000000
ptr_ring(size:1000) perf simple test for 100000000 times, took 2211409 us

Performance counter stats for './ptr_ring_test_opt -s 1000 -m 0 -N 100000000':

2211.70 msec task-clock # 1.000 CPUs utilized
24 context-switches # 0.011 K/sec
0 cpu-migrations # 0.000 K/sec
57 page-faults # 0.026 K/sec
5750136694 cycles # 2.600 GHz
16124176577 instructions # 2.80 insn per cycle
<not supported> branches
6553023 branch-misses

2.211968470 seconds time elapsed

2.210303000 seconds user
0.000000000 seconds sys

---------------patched ptr_ring.c end----------------------------------

>
>>