[Gluster-devel] Proposal for change regarding latency calculation

Fri Aug 1 09:54:46 UTC 2014

Vipul,

----- Original Message -----
> I would want tests of all the posix operations. Need a difference not just in
> throughput, but in max iops for the various ops.

The test that I suggested is incomplete. Like Joe has suggested,
we should see the effect of profiling on all FOPs. This could be done modifying
tests/basic/fops-sanity.t (present in glusterfs repo) a little bit.

This by itself could take longer than we might like it to from glusterfsiostat
perspective. How inconvenient would it be to make glusterfsiostat return an
error message to the user when profiling is disabled on the volume whose
statistics are queried? This could be the approach for the first cut.
Once we have settled that performance impact of enabling profiling by default is bearable, 
we could change glusterfsiostat to work seamlessly.

Does that make sense?

~KP

> 
> On 07/27/2014 08:27 AM, Vipul Nayyar wrote:
> 
> 
> 
> Hi
> 
> As guided by you, I performed the experiment regarding measurement of the
> effect of always enabled profiling. I performed two write tests, one with a
> 20 MB file and the other with a 730 MB file. Each file was written 20 times
> to the mounted volume after clearing the buffers on every iteration and the
> time taken measured with the time command. I ran the following bash script
> for this purpose.
> 
> i=1
> while [[ $i -lt 21 ]]; do
> sync && echo 3 > /proc/sys/vm/drop_caches
> path="/mnt/write_test"$i
> out=$( time cp /home/vipul/test.avi $path)
> i=$((i+1))
> done
> 
> Since the values at different times for writing the same file are quite
> varied, I plotted a graph using the obtained values(Y-axis represents
> seconds) which can be found attached. As you might see in these images,
> there is no clear pattern found in the variation of values obtained while
> writing.
> 
> So according to me, values in both the conditions are quite near to each
> other and equally capable of going quite high or low than the mean value and
> hence, there is no negative effect seen due to the change proposed. I hope
> someone else can shed more light on whether setting the option(always
> enabled profiling) really decreased the performance or not.
> 
> Regards
> Vipul Nayyar
> 
> 
> 
> On Wednesday, 16 July 2014 12:50 PM, Krishnan Parthasarathi
> <kparthas at redhat.com> wrote:
> 
> 
> Vipul,
> 
> 
> 
> 
> 
> Hello,
> 
> Following is a proposal for modifying the io profiling capability of the
> io-stats xlator. I recently sent in a patch(review.gluster.org/#/c/8244/)
> regarding that, which uses the already written latency related functions in
> io-stats to dump info through meta and added some more data containers which
> would track some more fops related info each time a request goes through
> io-stats. Currently, before the io-stats' custom latency functions can run,
> the measure_latency and count_fop_hits option should be enabled. I propose
> to remove these two options entirely from io-stats.
> 
> In order to track io performance, these options should be enabled all the
> time, or removed entirely, so that a record of io requests can be kept since
> mount time, else enabling these options only when it is required will not
> give you the average statistics over the whole period since the start. This
> is based on the methodology of Linux kernel itself, since it internally
> maintains the io statistics data structures all the time and presents it via
> /proc filesystem whenever required. Enabling of any option is not required,
> and the data available represents statistics since the boot time.
> 
> I would like to know the views over this, if having io-stats profiling info
> available all the time would be a good thing?
> Could you run the following experiment to measure the effect of profiling
> being enabled always?
> - Fix the I/O workload to be run.
> - Setup 1 (control group) : Run the fixed workload on a volume with both the
> profiling options NOT set.
> - Setup 2 : Run the (same) fixed workload on the same volume with the
> profiling options set.
> - In both setup, measure the latencies observed by the said workload. You
> could use time(1) command
> for a crude measurement.
> 
> This should allow us to make an informed decision on whether there is any
> performance effect
> when profiling is enabled on a volume by default.
> 
> 
> 
> 
> Apart from this, I was going over latency.c in libglusterfs, which does a
> fine job of maintaining latency info for every xlator and encountered an
> anomaly which I thought should be dealt with. The function
> gf_proc_dump_latency_info which dumps the latency array for the specified
> xlator consists of a last line which in the end flushes this array through
> memset after every dump. That means, you get different latency info every
> time you read the profile file in meta. I think, flushing the data structure
> after every dump is wrong since, you don't get overall stats since one
> enabled the option at the top of meta, and more importantly, multiple
> applications reading this file can get wrong info, since it gets cleared
> after one read only.
> Clearing of the statistics on every request sounds incorrect to me. Could you
> please send a patch to fix this?
> 
> thanks,
> Krish
> 
> 
> 
> 
> 
> If my reasons seem apt for you, I'll send a patch over for evaluation.
> 
> Regards
> Vipul Nayyar
> 
> 
> 
> 
> 
> 
> _______________________________________________
> Gluster-devel mailing list Gluster-devel at gluster.org
> http://supercolony.gluster.org/mailman/listinfo/gluster-devel
> 
> 
> _______________________________________________
> Gluster-devel mailing list
> Gluster-devel at gluster.org
> http://supercolony.gluster.org/mailman/listinfo/gluster-devel
> 