Why I/O test on LVM image (KVM) works suspicious compared to raw file image via varmail.f in Filebench benchmark?

Zhang Zhaoning <zzningxp@gmail.com> · Fri, 1 Mar 2013 15:19:58 +0800

Here I compare I/O performance on LVM image and raw file image (directly generated by Linux bash 'dd'), with the virtual machine hypervisor KVM, guest and host are both CentOS 6.2, and the guest has 1 CPU, 8G memory.
I use Filebench (filebench-1.4.9.1) as one of the benchmark, I use varmail.f oltp.f fileserver.f webserver.f (unmodified from the workload in the packet) to test the macro-benchmark.

The test of fileserver.f and webserver.f show that the LVM image and raw file image performs the similar result.
But the varmail.f shows that LVM outperforms than the raw file by several times, and the table below shows the data:

  fileserver.f (KB/s)
  oltp.f
(KB/s)
  varmail.f
(KB/s)
  webserver.f
(KB/s)

  rawfile
  17318.4
  76.8
  256
  8576

  lvm
  16921.6
  140.8
  780.8
  8448

Thus, I test varmail.f by changing different parameters, and the table below shows the data:

  lvm(a)
(KB/s)
  rawfile(b)
(KB/s)
  a/b

  varmail.iosize-.f
  742.4
  179.2
  4.142857

  varmail.iosize+.f
  742.4
  179.2
  4.142857

  varmail.meanappendsize-.f
  51.2
  25.6
  2

  varmail.meanappendsize+.f
  2278.4
  985.6
  2.311688

  varmail.meandirwidth-.f
  729.6
  179.2
  4.071429

  varmail.meandirwidth+.f
  742.4
  166.4
  4.461538

  varmail.meanfilesize-.f
  729.6
  179.2
  4.071429

  varmail.meanfilesize+.f
  742.4
  179.2
  4.142857

  varmail.nfiles-.f
  819.2
  192
  4.266667

  varmail.nfiles+.f
  640
  179.2
  3.571429

  varmail.nthreads-.f
  486.4
  153.6
  3.166667

  varmail.nthreads+.f
  1190.4
  230.4
  5.166667

  varmail.nofsync.f
  6259.2
  6272
  0.997959

varmail.nofsync.f is remove these two lines:
    flowop fsync name=fsyncfile2,fd=1
    flowop fsync name=fsyncfile3,fd=1

So, the initiator is fsync, which is not existed in fileserver.f and webserver.f, and oltp.f has a dsync.
I think the function of the fsync is to force data from OS memory to disk, but how does it influence the LVM virtual I/O performance?

Could anybody told me what cause this?

Thanks very much.

-- 
Thank you!

Brian Zhaoning Zhang

_______________________________________________
linux-lvm mailing list
linux-lvm@redhat.com
https://www.redhat.com/mailman/listinfo/linux-lvm
read the LVM HOW-TO at http://tldp.org/HOWTO/LVM-HOWTO/

	fileserver.f (KB/s)	oltp.f (KB/s)	varmail.f (KB/s)	webserver.f (KB/s)
rawfile	17318.4	76.8	256	8576
lvm	16921.6	140.8	780.8	8448

	lvm(a) (KB/s)	rawfile(b) (KB/s)	a/b
varmail.iosize-.f	742.4	179.2	4.142857
varmail.iosize+.f	742.4	179.2	4.142857
varmail.meanappendsize-.f	51.2	25.6	2
varmail.meanappendsize+.f	2278.4	985.6	2.311688
varmail.meandirwidth-.f	729.6	179.2	4.071429
varmail.meandirwidth+.f	742.4	166.4	4.461538
varmail.meanfilesize-.f	729.6	179.2	4.071429
varmail.meanfilesize+.f	742.4	179.2	4.142857
varmail.nfiles-.f	819.2	192	4.266667
varmail.nfiles+.f	640	179.2	3.571429
varmail.nthreads-.f	486.4	153.6	3.166667
varmail.nthreads+.f	1190.4	230.4	5.166667
varmail.nofsync.f	6259.2	6272	0.997959