[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Freeipa-devel] [PATCH] test speedup patch



On 01/19/2011 04:17 AM, Rob Crittenden wrote:
Rob Crittenden wrote:
Rob Crittenden wrote:
Attached is a rough cut of a patch to try to speed up the cli a little
bit. Basically in production mode it will skip some things during
initialization.

My concept is that we develop in mode != production and release in mode
== production.

I managed to knock a second or so off time to do a user-show on average.

There may be some other things we can do to speed things up, I'm still
looking. Some feedback on the approach would be appreciated.

Note that I've completely ruled out SSL/Negotiate. I did my testing on
lite-server which doesn't use SSL or Negotiate and it was STILL taking
on average 3-4+ seconds per command. The server side was consistently
taking < 1 second to complete.

rob

oh, and the patch.

I ran a couple of moderate tests this evening that executed 42 separate
operations like add, delete, and managing group membership. I ran this
10 times each on 2 identical VMs, one with a bit older code and one with
this patch then averaged the times.

With the patch the average was 1.3 seconds per operation, without 2.6. A
50% improvement is more than I expected, I saw a 33% improvement on
individual runs. I'll keep at it but this seems promising. I was also a
bit surprised that the average time without the patch was so low, I was
expecting something over 3 seconds.

Specifically what this patch does is it avoids doing some
self-validation. There is some amount of risk that the framework could
blow up but in a deployed situation I think the risk is rather low.

A side-effect of the API tester makeapi is that it loads the framework.
We can force it to be run in production mode so the product shouldn't be
buildable if it has inconsistencies.

rob


I find it hard to believe this patch causes such a big improvement in performance. Especially the parts skipping asserts, that shouldn't be significantly slower than your average ifs. Instance locking shouldn't be a time consuming operation either.

Bypassing check routines for parameter namespaces might provide a performance boost as it is called for every single plugin we have (~250). On the other hand, it is only used for positional arguments and most plugins only have 1 or 2 of those.

Personally, I would do some more tests on a single machine, because there's no guarantee, that two VMs with an identical image have the same performance.

If it really provides a significant improvement, than it's awesome, because I like the philosophy of this patch. It removes self-checking and instance locking, that is completely useless in a production environment and kind of limiting in non-production. I think there's more places likes this in the framework.

Long story short:
It's improbable, but not impossible, for the changes introduced by this patch to cause such a big performance improvement. Even if it doesn't, the patch is still good.

Pavel


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]