helicoterus-elih@yahoo.it
2009-May-24 12:25 UTC
[Xen-users] Difference performances between HVM and PV guest in hard disc reading
I perform some tests to get hard disc performance differences between HVM domU
and a paravirtualized one (both Debian Lenny).
I used hdparm ( option -t and -T) to test cache of disc and disc sequential
read.
Syntax is:
hdparm -t|-T /dev/hda1
hdparm -t for hvm returns about 30 MB/sec, wheras hdparm -t for PV return about
180 MB/sec.
hdparm -T for hvm returns about 1247 MB/sec, wheras hdparm -T for PV return
about 1183 MB/sec.
Are this test true? What about this differences? Why are hard disc performances
so different?
Thanks
_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users
Fajar A. Nugraha
2009-May-24 22:36 UTC
Re: [Xen-users] Difference performances between HVM and PV guest in hard disc reading
On Sun, May 24, 2009 at 7:25 PM, <helicoterus-elih@yahoo.it> wrote:> > I perform some tests to get hard disc performance differences between HVM domU and a paravirtualized one (both Debian Lenny).first of all, make sure you do NOT use file:/ in domU config. At least use tap:aio:> I used hdparm ( option -t and -T) to test cache of disc and disc sequential read. > Syntax is: > hdparm -t|-T /dev/hda1 > > hdparm -t for hvm returns about 30 MB/sec, wheras hdparm -t for PV return about 180 MB/sec.This is normal. HVM by default should have MUCH lower disk throughput. That''s why HVM domUs need PV drivers for both Linux (included by default on RHEL/Centos 5.3, not available yet for newer kernels) and Windows (like GPLPV).> hdparm -T for hvm returns about 1247 MB/sec, wheras hdparm -T for PV return about 1183 MB/sec.-T is cached result. Repeat this several times and you should find similar output for both PV and HVM. -- Fajar _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
Heli
2009-May-25 18:50 UTC
Re: [Xen-users] Difference performances between HVM and PV guest in hard disc reading
Fajar A. Nugraha ha scritto:> On Sun, May 24, 2009 at 7:25 PM, <helicoterus-elih@yahoo.it> wrote: >> I perform some tests to get hard disc performance differences between HVM domU and a paravirtualized one (both Debian Lenny). > > first of all, make sure you do NOT use file:/ in domU config. At least > use tap:aio: > >> I used hdparm ( option -t and -T) to test cache of disc and disc sequential read. >> Syntax is: >> hdparm -t|-T /dev/hda1 >> >> hdparm -t for hvm returns about 30 MB/sec, wheras hdparm -t for PV return about 180 MB/sec. > > This is normal. HVM by default should have MUCH lower disk throughput. > That''s why HVM domUs need PV drivers for both Linux (included by > default on RHEL/Centos 5.3, not available yet for newer kernels) and > Windows (like GPLPV).Where I can find this drivers for debian? I would like to test a debian HVM virtual machine with these "magic" drivers. I only saw windows ones. Thanks Chiacchiera con i tuoi amici in tempo reale! http://it.yahoo.com/mail_it/foot/*http://it.messenger.yahoo.com _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
Fajar A. Nugraha
2009-May-26 02:18 UTC
Re: [Xen-users] Difference performances between HVM and PV guest in hard disc reading
On Tue, May 26, 2009 at 1:50 AM, Heli <helicoterus-elih@yahoo.it> wrote:> Where I can find this drivers for debian? > I would like to test a debian HVM virtual machine with these "magic" > drivers.I haven''t find one that works. In the future virtio drivers might work, but at this moment last time I check PV drivers for Linux HVM guest only work on kernel 2.6.18. If you want to use it for production purposes, its easy enough to convert HVM guest to PV. Here''s an example for Ubuntu and RHEL: http://lists.xensource.com/archives/html/xen-users/2009-05/msg00536.html If you only want to do some testing, you should be able to import RHEL/Centos 5.3''s -xen kernel RPM to debian with alien. I tested this some time ago. You need some modifications on /etc/initramfs-tools/modules (does Debian use the same file?) to add some modules to initrd. Another option is to use kernel from xen.org (2.6.18.8) and compile the drivers yourself (the source is on xen tarball) -- Fajar _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
Heli
2009-May-30 17:20 UTC
Re: [Xen-users] Difference performances between HVM and PV guest in hard disc reading
Fajar A. Nugraha wrote:> On Sun, May 24, 2009 at 7:25 PM, <helicoterus-elih@yahoo.it> wrote: >> I perform some tests to get hard disc performance differences between HVM domU and a paravirtualized one (both Debian Lenny). > >> I used hdparm ( option -t and -T) to test cache of disc and disc sequential read. >> Syntax is: >> hdparm -t|-T /dev/hda1 >> >> hdparm -t for hvm returns about 30 MB/sec, wheras hdparm -t for PV return about 180 MB/sec. > > This is normal. HVM by default should have MUCH lower disk throughput. > That''s why HVM domUs need PV drivers for both Linux (included by > default on RHEL/Centos 5.3, not available yet for newer kernels) and > Windows (like GPLPV). > >> hdparm -T for hvm returns about 1247 MB/sec, wheras hdparm -T for PV return about 1183 MB/sec. > > -T is cached result. Repeat this several times and you should find > similar output for both PV and HVM. >How is it possible that a PV domU returns 180 MB/sec if on a normal linux hdparm returns 45 MB/sec? So PV is very crazy!!!Chiacchiera con i tuoi amici in tempo reale! http://it.yahoo.com/mail_it/foot/*http://it.messenger.yahoo.com _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
Javier Guerra
2009-May-31 01:02 UTC
Re: [Xen-users] Difference performances between HVM and PV guest in hard disc reading
Heli wrote:> How is it possible that a PV domU returns 180 MB/sec if on a normal > linux hdparm returns 45 MB/sec?easy, if the ''block device'' is in fact a file on Dom0, with a lot more RAM available for cache but as soon as you do real load, that cache becomes a nuisianse. you get much better performance if you turn the cache off (with tap:) or use a block device (with phy:), even if the benchmarks drop down to realistic levels -- Javier _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users
Fajar A. Nugraha
2009-May-31 01:03 UTC
Re: [Xen-users] Difference performances between HVM and PV guest in hard disc reading
On Sun, May 31, 2009 at 12:20 AM, Heli <helicoterus-elih@yahoo.it> wrote:> How is it possible that a PV domU returns 180 MB/sec if on a normal linux > hdparm returns 45 MB/sec?Did you read my first email? "first of all, make sure you do NOT use file:/ in domU config." I wrote that for a reason, you know. :) Short story, dom0 caching effect can skew performance benchmark results. -- Fajar _______________________________________________ Xen-users mailing list Xen-users@lists.xensource.com http://lists.xensource.com/xen-users