[Gluster-devel] [Gluster-users] What's the impact of enabling the profiler?

Joe Julian joe at julianfamily.org
Tue Jul 22 06:26:23 UTC 2014


On 07/21/2014 11:20 PM, Pranith Kumar Karampuri wrote:
>
> On 07/22/2014 11:39 AM, Joe Julian wrote:
>>
>> On 07/17/2014 07:30 PM, Pranith Kumar Karampuri wrote:
>>>
>>> On 07/18/2014 03:05 AM, Joe Julian wrote:
>>>> What impact, if any, does starting profiling (gluster volume 
>>>> profile $vol start) have on performance?
>>> Joe,
>>>     According to the code the only extra things it does is calling 
>>> gettimeofday() call at the beginning and end of the FOP to calculate 
>>> latency, increment some variables. So I guess not much?
>>>
>>
>> So far so good. Is the only way to clear the stats to restart the brick?
> I think when the feature is initially proposed we wanted two things
> 1) cumulative stats
> 2) Interval stats
>
> Interval stats get cleared whenever 'gluster volume profile <volname> 
> info' is executed (Although it starts counting the next set of fops 
> that happen after this command execution). But there is no way to 
> clear the cumulative stats. It would be nice if you could give some 
> feedback about what you liked/what you think should change to make 
> better use of it. So I am guessing there wasn't big performance hit?
>
> Pranith
No noticeable performance hit, no.

I'm writing a whitepaper for the best practices for OpenStack on 
GlusterFS so I needed some idea how qemu actually uses the filesystem. 
What the operations are so I can look at not only the best ways to tune 
for that use, but how to build the systems around that.

At this point, I'm just collecting data. TBH, I hadn't noticed the 
interval data. That should be perfect for this. I'll poll it in XML and 
run the numbers in a few days.


More information about the Gluster-devel mailing list