Significant performance degradation on the SDK server with 1.5.3

87 views
Skip to first unread message

PK

unread,
Sep 6, 2011, 11:11:53 PM9/6/11
to google-a...@googlegroups.com
This past weekend I noticed that my tests take about 50% longer to run than what they used to. Tonight, I bisected my code and nailed this down to just changing one thing: using 1.5.3 vs using 1.5.2 Here is typical output for one of the tests:

$ time ....
GAE SDK 1.5.2:                     GAE SDK 1.5.3:
real    0m50.233s                   1m43.665s
user    0m20.450s                  1m17.560s
sys     0m3.080s                    0m3.000s

All my tests use a combination of selenium and remote APIs, the environment is python, using the original datastore (no mysql), the tests run on ubuntu 10.04.3.

Anybody who noticed similar behavior? Any explanation from the SDK team, do you have any internal benchmarks that you use before each release?

Thanks,
PK

PK

unread,
Nov 16, 2011, 9:15:22 PM11/16/11
to google-a...@googlegroups.com
Today I run the test again on a high end MacBook Pro 10.7.2. Python is 2.5.5.  Exactly the same code base, I just change the SDK and run the test again to get the numbers!!!

Here are the results, it seems to be getting worse:

SDK:            1.5.2                  1.5.5                1.6.0
real              0m41.496s       1m33.293s        1m55.807s
user             0m13.629s       1m15.434s        1m26.100s
sys              0m4.530s         0m7.297s          0m9.128s        

PK

unread,
Nov 17, 2011, 3:32:42 PM11/17/11
to google-a...@googlegroups.com
R.S. in another thread was kind enough to point me to the issue and workaround reported in issue 6005. I rerun the test using the workaround. It indeed is responsible for the majority of the performance regression i have been experiencing:

SDK:    1.6.0
real       0m53
user      0m23
sys       0m6

PK

unread,
Dec 6, 2011, 3:08:51 PM12/6/11
to google-a...@googlegroups.com
Wow!! I just benchmarked the same test case with 1.6.1 prerelease without any workarounds and am happy to report that that performance issues have been addressed:

real 0m35.221s
user 0m16.330s
sys 0m2.460s

Thanks for taking care of the issues!!!

PK
Reply all
Reply to author
Forward
0 new messages