On Wednesday, March 25, 2020 3:25 PM, Kirk Jamison wrote:
> As for the performance and how it affects the read-only workloads.
> Using pgbench, I've kept the overload to a minimum, less than 1%.
> I'll post follow-up results.
Here's the follow-up results.
I executed the similar tests from top of the thread.
I hope the performance test results shown below would suffice.
If not, I'd appreciate any feedback with regards to test or the patch itself.
A. VACUUM execution + Failover test
- 100GB shared_buffers
1. 1000 tables (18MB)
1.1. Execution Time
- [MASTER] 77755.218 ms (01:17.755)
- [PATCH] Execution Time: 2147.914 ms (00:02.148)
1.2. Failover Time (Recovery WAL Replay):
- [MASTER] 01:37.084 (1 min 37.884 s)
- [PATCH] 1627 ms (1.627 s)
2. 10000 tables (110MB)
2.1. Execution Time
- [MASTER] 844174.572 ms (14:04.175) ~14 min 4.175 s
- [PATCH] 75678.559 ms (01:15.679) ~1 min 15.679 s
2.2. Failover Time:
- [MASTER] est. 14 min++
(I didn't measure anymore because recovery takes
as much as the execution time.)
- [PATCH] 01:25.559 (1 min 25.559 s)
Significant performance results for VACUUM.
B. TPS Regression for READ-ONLY workload
(PREPARED QUERY MODE, NO VACUUM)
# [16 Clients]
- pgbench -n -S -j 16 -c 16 -M prepared -T 60 cachetest
|shbuf |Master |Patch |% reg |
|----------|--------------|---------------|----------|
|128MB| 77,416.76 | 77,162.78 |0.33% |
|1GB | 81,941.30 | 81,812.05 |0.16% |
|2GB | 84,273.69 | 84,356.38 |-0.10%|
|100GB| 83,807.30 | 83,924.68 |-0.14%|
# [1 Client]
- pgbench -n -S -c 1 -M prepared -T 60 cachetest
|shbuf |Master |Patch |% reg |
|----------|--------------|---------------|----------|
|128MB| 12,044.54 | 12,037.13 |0.06% |
|1GB | 12,736.57 | 12,774.77 |-0.30%|
|2GB | 12,948.98 | 13,159.90 |-1.63%|
|100GB| 12,982.98 | 13,064.04 |-0.62%|
Both were run for 10 times and average tps and % regression are
shown above. At some point only minimal overload was caused by
the patch. As for other cases, it has higher tps compared to master.
If it does not make it this CF, I hope to receive feedback in the future
on how to proceed. Thanks in advance!
Regards,
Kirk Jamison
