Xen users - Sep 2007 - Physical Disk VBD vs File VBD Performance

Hi Everyone, disk I/O performance question here.

I have system that uses RAID5 for the the dom0 and I have to domUs
running.  One dom1 uses a file based VBD and dom2 uses another RAID0
array *with one disk*.  I know you say "wtf, one disk???", but
I''ll
explain later.

When doing 2GB write tests with dd command (see below) on the dom0
I get about 64MB/s to the disk.  Now when I write the same dd command
out on dom2 (physical VBD, RAID0) I get about 30MB/s to the disk.  I then
asked myself if this made sense.  Correct me if I''m wrong, when writing
to RAID5 the data gets striped up and your write performance increases.
Now if I write to the dom2 it has to write the whole data to one spindle.

`dd if=/dev/zero of=file_2gb_1024k_blocks bs=1024k count=2000`

So for shucks n grins, I tested the dom1 (file VBD, RAID5/dom0)
write performance and nearly had a stroke when I saw that it reached about
54MB/s write speed.  I understand that writing on dom1 is faster because of
the RAID5 on dom0.

Granted I''m only writing one file to the disk, but is it better in
overall
performance, considering multiple concurrent file reads/writes that dom2 will
experience, to use the physical VBD?  Will the file VBD degrade in performance
earlier than the physical VBD under the same heavy load earlier?

I''m using dom1 as a web server and dom2 as a database server.

About why I''m using one spindle on for the RAID0 array is that I can
only
afford 1 drive right now to get started and Dell has assured me that I can
add two more drives to form a RAID5 array - ultimately, it is my goal.

dom0 = Dual Proc, Dual core Woodcrest,6GB (total), RHEL 4.4, RAID5/3 ultra 320
10K rpm.
dom1 = Centos 5, file VBD, 2GB.
dom2 = Centos 5, physical VBD RAID0/1 ultra 320 10K rpm, 2GB.

Thanks for your time and knowledge,
Matt

_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

Matias wrote:
> Correct me if I''m wrong, when writing
> to RAID5 the data gets striped up and your write performance increases.
> Now if I write to the dom2 it has to write the whole data to one spindle.
RAID5 frequently causes performance losses, not gains, although with 
high performance controllers and disks you can see performance 
improvements over a raw disk.  The problem is that the data isn''t just 
striped like in a RAID-0 set, the data has to have parity information 
calculated and written out as well, and all disks get hit with write 
requests at the same time when writing out data.  On reads (in a 
non-degraded set), you may see performance increases because of the 
striped nature of the data, but no guarantees.
-- 
Michael T. Babcock
http://mikebabcock.ca


_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

"Michael T. Babcock" <mike@mikebabcock.ca> writes:
> Matias wrote:
>> Correct me if I''m wrong, when writing
>> to RAID5 the data gets striped up and your write performance increases.
>> Now if I write to the dom2 it has to write the whole data to one
spindle.
>
> RAID5 frequently causes performance losses, not gains, although with
> high performance controllers and disks you can see performance
> improvements over a raw disk.  The problem is that the data isn''t
just
> striped like in a RAID-0 set, the data has to have parity information
> calculated and written out as well, and all disks get hit with write
> requests at the same time when writing out data.  On reads (in a
> non-degraded set), you may see performance increases because of the
> striped nature of the data, but no guarantees.
Note that you only gain anything when writing a full stripe. Updating
a stripe entails reading the parity block, updating it, writing
it. That costs you more than the speedup of striped data.

Databases are know to cause small random writes killing your raid5
performance while filesystem try to keep data continious and don''t
suffer (much).

MfG
        Goswin

_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users