[ovs-dev] [PATCH 8/8] lib/classifier: Use cmap.

[Jarno Rajahalme]

I think I found the reason for the slowness of cmap on XenServer.

I was debugging on a XenServer tonight and found out that the configure test for GCC atomics had apparently failed, as I noticed actual lock and unlock operations on gdb disassembly, and then by the fact the changes I made to the ovs-atomic-gcc4+.h did not trigger any recompilations with “make” (so there was no dependency on it).

If the configure check is correct, then there must be something wrong with the XenServer setup?

  Jarno

On Jun 25, 2014, at 11:09 AM, Ben Pfaff <blp at nicira.com> wrote:

> I am really surprised to hear that it takes over a minute.  On my
> workstation (which is a few years old), "ovstest test-cmap check 1"
> only takes 6.5 seconds.  Do you have any idea what's different on
> XenServer?  It might be the GCC version, which version is it using?
> 
> On Wed, Jun 25, 2014 at 09:40:12AM -0700, Gurucharan Shetty wrote:
>> I had a quick comment. I noticed while running unit tests on Xenserver
>> that the cuckoo hash unit test was very slow. It takes a little more
>> than one minute to complete. I wonder whether using more cmap would
>> hurt Xenserver's performance?
>> 
>> On Wed, Jun 25, 2014 at 4:19 AM, Jarno Rajahalme <jrajahalme at nicira.com> wrote:
>>> Ben,
>>> 
>>> Thanks for the reviews!
>>> 
>>> I have addressed the (minor) comments on the series, will push once the vector v2 is ack???ed. I???ll keep this last patch as the first in the classifier RCU series.
>>> 
>>>  Jarno
>>> 
>>> On Jun 13, 2014, at 10:37 AM, Ben Pfaff <blp at nicira.com> wrote:
>>> 
>>>> On Mon, Jun 09, 2014 at 11:53:55AM -0700, Jarno Rajahalme wrote:
>>>>> Use cmap instead of hmap & hindex in classifier.  Performance impact
>>>>> with current locking strategy is not yet tested.  Later patches will
>>>>> introduce RCU into the classifer.
>>>> 
>>>> into the "classifier".
>>>> 
>>>>> Signed-off-by: Jarno Rajahalme <jrajahalme at nicira.com>
>>>> 
>>>> As long as this passes the tests (I assume you ran them) it seems OK
>>>> to me.  The tricky review will be when we start actually making the
>>>> classifier RCU.  (I don't know if it makes sense to push this patch
>>>> before then.  It might be better as the first patch of a series that
>>>> makes the classifier RCU.  But I'll leave that decision to you.)
>>>> 
>>>> Acked-by: Ben Pfaff <blp at nicira.com>
>>> 
>>> _______________________________________________
>>> dev mailing list
>>> dev at openvswitch.org
>>> http://openvswitch.org/mailman/listinfo/dev