On Thu, Apr 14, 2016 at 5:35 AM, Andres Freund <andres@anarazel.de> wrote:
On 2016-04-14 07:59:07 +0530, Amit Kapila wrote: > What you want to see by prewarming?
Prewarming appears to greatly reduce the per-run variance on that machine, making it a lot easier to get meaningful results. Thus it'd make it easier to compare pre/post padding numbers. > Will it have safe effect, if the tests are run for 10 or 15 mins > rather than 5 mins?
s/safe/same/? If so, no, I doesn't look that way. The order of buffers appears to play a large role; and it won't change in a memory resident workload over one run.
I've tried to run read-only benchmark of pad_pgxact_v1.patch on 4x18 Intel machine. The results are following.
clients no-pad pad
1 12997 13381
2 26728 25645
4 52539 51738
8 103785 102337
10 132606 126174
20 255844 252143
30 371359 357629
40 450429 443053
50 497705 488250
60 564385 564877
70 718860 725149
80 934170 931218
90 1152961 1146498
100 1240055 1300369
110 1207727 1375706
120 1167681 1417032
130 1120891 1448408
140 1085904 1449027
150 1022160 1437545
160 976487 1441720
170 978120 1435848
180 953843 1414925
snapshot_too_old patch was reverted in the both cases.
On high number of clients padding gives very significant benefit. However, on low number of clients there is small regression. I think this regression could be caused by alignment of other data structures.
------ Alexander Korotkov Postgres Professional: http://www.postgrespro.com The Russian Postgres Company