|
I believe that this is simply caused by the merge of MySQL 5.7.9 into MariaDB 10.2.2.
Because Jan Lindström did that, I am assigning this to him for now.
|
|
I am assigning this back to me.
I wonder how much we have improved since those days. I think that already 10.5.12 or 10.6.4 should perform considerably better, and 10.6.5 includes some more improvements.
|
|
Today, I observed 130ktps at 16 threads on 10.6 when testing the Linux kernel io_uring hang (MDEV-26674). But the hardware is newer and the workload is different, so this is not directly comparable. Besides, it was with an insanely small redo log, to trigger very frequent page writes in order to test the kernel bug. During furious page flushing (frequent checkpoints due to the intentional misconfiguration), the throughput dropped to the claimed 10.2 level (60ktps).
|
|
Things got no better in recent releases:
| thread count |
10.2.1 QPS |
10.2.2 QPS |
10.6.5 QPS |
| 1 |
7233.9 |
4622.5 |
4874.7 |
| 8 |
43185 |
32226 |
32260 |
| 16 |
75237 |
61746 |
59867 |
| 32 |
130255 |
114321 |
98753 |
| 64 |
168282 |
173745 |
155256 |
| 128 |
179679 |
205590 |
191656 |
| 256 |
179779 |
209018 |
203569 |
|
|
Attached a summary of all InnoDB-related benchmarks from the regression suite for 10.2.1, 10.2.2 and 10.6.5 - MDEV-13190.pdf
|
|
I have rerun the benchmarks using the current kernel (with mitigations against SPECTRE & co) and have also recompiled everything, so using the same compiler and libraries.
I still see an increase in latency and accordingly a decrease in performance at low thread counts for 10.2.2 (vs. 10.2.1). 10.6.5 looks quite good in direct comparison. For read-only it is better du to faster collation algorithm. For writes 10.6 is just better.
Since the regression vanishes at higher load, it seems like the extra time is spent waiting in the kernel with the user threads suspended.
Attached: sysbench_de.pdf (direct execution), sysbench_ps.pdf (prepared statements)
|
|
Thank you. Some regression for updates at low thread counts is still present. I hope that applying https://www.brendangregg.com/offcpuanalysis.html in a similar way as in MDEV-26004 could identify the bottleneck.
The regression for point selects happens at any thread count. I wonder if using the READ UNCOMMITTED isolation level would make it go away. Possibly MDEV-21423 could address this.
|
|
Tested with transaction-isolation = 'READ-UNCOMMITTED'. For point-selects the performance didn't change at all, for read-only it became actually slower.
Attached: sysbench_read_uncommitted.pdf
|
MDEV-13190.pdf