[Gluster-devel] [RFC PATCH v0 0/1] Zero copy readv

[Bharata B Rao]

On Wed, Mar 6, 2013 at 5:24 PM, Bharata B Rao <bharata.rao at gmail.com> wrote:
> On Tue, Mar 5, 2013 at 9:41 PM, Anand Avati <anand.avati at gmail.com> wrote:
>>
>> Do you have comparison of the %cpu util with and without zero-copy? cpu util
>> is probably be an important parameter of comparison.
>
> Right, I don't have the numbers. BTW, I am seeing that the results are
> a bit fragile at the moment and I am trying to get the basic read
> numbers by directly reading from GlusterFS volume instead of QEMU and
> that should give us a better comparison.

The CPU utilization from perf stat looks like this for FIO on QEMU guest:

W/o zero copy
===========
83800.891583 task-clock    # 0.258 CPUs utilized
120,843,564,441 cycles      # 1.442 GHz
324.536693239 seconds time elapsed

Zero copy
=======
54320.89935 task-clock     # 0.177 CPUs utilized
111,921,816,282 cycles     # 2.060 GHz
307.138280851 seconds time elapsed

The following numbers are for sequential read (4096 bytes at a time)
of a 3GB file residing on GlusterFS volume backed by tmpfs.

W/o zero copy
============
269182.305852 task-clock                #    1.325 CPUs utilized
809,573,871,914 cycles                    #    3.008 GHz
      [83.37%]
203.089330302 seconds time elapsed

Zero copy
========
269120.040663 task-clock                #    1.315 CPUs utilized
806,918,650,105 cycles                    #    2.998 GHz
      [83.29%]
204.702692073 seconds time elapsed

As seen from the above numbers, I don't see any difference. May be I
should test by having the entire VM image on GlusterFS volume backed
by tmpfs to eliminate IO bottleneck.

Regards,
Bharata.
-- 
http://raobharata.wordpress.com/