Re: [CentOS] Really bad KVM disk performance
On Mon, Feb 20, 2012 at 02:26, Bob Puff b...@nleaudio.com wrote: Hi Gang, [...] On my machine's Centos 5.7 x32 guest install: # hdparm -tT /dev/hda /dev/hda: Timing cached reads: 1864 MB in 2.16 seconds = 863.87 MB/sec Timing buffered disk reads: 358 MB in 3.08 seconds = 116.17 MB/sec cached reads is a measure of linux buffer bandwitdth, not disk performance!!. It should be several thousands... In a real machine, it could be a motherboard problem... -- Marcelo ¿No será acaso que ésta vida moderna está teniendo más de moderna que de vida? (Mafalda) ___ CentOS mailing list CentOS@centos.org http://lists.centos.org/mailman/listinfo/centos
Re: [CentOS] Really bad KVM disk performance
On 02/19/2012 09:26 PM, Bob Puff wrote: Immediately when I was installing stuff, I could tell this new system I just built was not nearly as fast as the first one. I ran some CPU and disk benchmarking programs, and saw that while the CPU stuff tested similarly, the disk thruput was much different... Down-right poor in one of the guests! If you use Red Hat's virt-install to set up a guest, you can specify the --os-type and --os-variant. If you specify the a variant that supports virtualized IO (such as type: linux and variant: rhel5.4 or rhel 6), the guest's disk and network IO will have much better throughput. You'll want to look for those options in whatever front-end you're using, or use virt-install. ___ CentOS mailing list CentOS@centos.org http://lists.centos.org/mailman/listinfo/centos
[CentOS] Really bad KVM disk performance
Hi Gang,
I recently rented a server at a datacenter with Centos 5.7 X64, Q9550
Processor, 8GB Ram, and dual 250GB SATA HDs (with 16mb cache). They had
loaded it with KVM, and installed a 30-day trial of Virtualizor as the
front-end for KVM.
I was so impressed with how fasts the guests ran that I want to build a few of
these machines for myself. I just installed one: same Q9550 processor, 4GB
ram, and dual 250GB SATA HDs (with 32mb cache). I installed Centos 6.2 X64,
and installed Webmin's Cloudmin as the front-end.
Immediately when I was installing stuff, I could tell this new system I just
built was not nearly as fast as the first one. I ran some CPU and disk
benchmarking programs, and saw that while the CPU stuff tested similarly, the
disk thruput was much different... Down-right poor in one of the guests!
On both systems, /dev/md2 is a LVM reserved exclusively for KVM guests. So
each guest is running in its own logical volume, in software raid.
Thinking there may be something wrong with the HDs, I ran Bonnie (
http://www.coker.com.au/bonnie++/ ) and compared both host machines. They
tested fairly similar (within 10%). Yet comparing their guests is like night
and day. Example:
On good machine's Centos 5.7 x32 guest install:
# hdparm -tT /dev/hda
/dev/hda:
Timing cached reads: 26760 MB in 1.99 seconds = 13417.10 MB/sec
Timing buffered disk reads: 388 MB in 3.01 seconds = 128.86 MB/sec
On my machine's Centos 5.7 x32 guest install:
# hdparm -tT /dev/hda
/dev/hda:
Timing cached reads: 1864 MB in 2.16 seconds = 863.87 MB/sec
Timing buffered disk reads: 358 MB in 3.08 seconds = 116.17 MB/sec
On one of my machine's Mandrake 8.2 x32 guest install:
# hdparm -tT /dev/hda
/dev/hda:
Timing buffer-cache reads: 27000 MB in 2.00 seconds = 13500.00 MB/sec
Timing buffered disk reads: 12 MB in 3.66 seconds = 3.28 MB/sec
On that system, the hdparm's -i output shows:
# hdparm -i /dev/hda
/dev/hda:
Model=QEMU HARDDISK, FwRev=0.12.1, SerialNo=QM1
Config={ Fixed }
RawCHS=16383/16/63, TrkSize=32256, SectSize=512, ECCbytes=4
BuffType=DualPortCache, BuffSize=256kB, MaxMultSect=16, MultSect=16
CurCHS=16383/16/63, CurSects=16514064, LBA=yes, LBAsects=73400320
IORDY=yes, tPIO={min:120,w/IORDY:120}, tDMA={min:120,rec:120}
PIO modes: pio0 pio1 pio2
DMA modes: sdma0 sdma1 sdma2 mdma0 mdma1 mdma2
UDMA modes: udma0 udma1 udma2 udma3 udma4 *udma5
AdvancedPM=no
Drive conforms to: ATA/ATAPI-5 published, ANSI NCITS 340-2000:
* signifies the current active mode
The bonnie numbers show for sequential output:
Good Machine Host: 76,857K/Sec
My Machine Host: 72,561K/Sec
Good Machine Centos 5.7 Guest: 66,266K/sec
My Machine Centos 5.7 Guest: 20,623K/sec
My machine Mandrake Guest: 1,365K/sec
Where should I look? I realize I do have two different front-ends to KVM, and
perhaps they are passing different parameters to it. I am also running the
KVM from Centos 6.2 on my machine, vs the other server is running on 5.7, but
I would have thought that newer is better. Also note that my hard drives
have a larger cache.
On a side note, I'm not thrilled with the Virtualizor's tech support, but the
product seems easy to use, once it actually works. Cloudmin seems to be
buggy, and not let you do things like change cd images on the fly, access the
console before the machine fully boots (!)... Any suggestions on other,
preferably open-source options? I'm a definite newbie to this virtualization
stuff.
Bob
___
CentOS mailing list
CentOS@centos.org
http://lists.centos.org/mailman/listinfo/centos
Re: [CentOS] Really bad KVM disk performance
How much did you paid for this?
On 20/02/2012 07:26 AM, Bob Puff wrote:
Hi Gang,
I recently rented a server at a datacenter with Centos 5.7 X64, Q9550
Processor, 8GB Ram, and dual 250GB SATA HDs (with 16mb cache). They had
loaded it with KVM, and installed a 30-day trial of Virtualizor as the
front-end for KVM.
I was so impressed with how fasts the guests ran that I want to build a few of
these machines for myself. I just installed one: same Q9550 processor, 4GB
ram, and dual 250GB SATA HDs (with 32mb cache). I installed Centos 6.2 X64,
and installed Webmin's Cloudmin as the front-end.
Immediately when I was installing stuff, I could tell this new system I just
built was not nearly as fast as the first one. I ran some CPU and disk
benchmarking programs, and saw that while the CPU stuff tested similarly, the
disk thruput was much different... Down-right poor in one of the guests!
On both systems, /dev/md2 is a LVM reserved exclusively for KVM guests. So
each guest is running in its own logical volume, in software raid.
Thinking there may be something wrong with the HDs, I ran Bonnie (
http://www.coker.com.au/bonnie++/ ) and compared both host machines. They
tested fairly similar (within 10%). Yet comparing their guests is like night
and day. Example:
On good machine's Centos 5.7 x32 guest install:
# hdparm -tT /dev/hda
/dev/hda:
Timing cached reads: 26760 MB in 1.99 seconds = 13417.10 MB/sec
Timing buffered disk reads: 388 MB in 3.01 seconds = 128.86 MB/sec
On my machine's Centos 5.7 x32 guest install:
# hdparm -tT /dev/hda
/dev/hda:
Timing cached reads: 1864 MB in 2.16 seconds = 863.87 MB/sec
Timing buffered disk reads: 358 MB in 3.08 seconds = 116.17 MB/sec
On one of my machine's Mandrake 8.2 x32 guest install:
# hdparm -tT /dev/hda
/dev/hda:
Timing buffer-cache reads: 27000 MB in 2.00 seconds = 13500.00 MB/sec
Timing buffered disk reads: 12 MB in 3.66 seconds = 3.28 MB/sec
On that system, the hdparm's -i output shows:
# hdparm -i /dev/hda
/dev/hda:
Model=QEMU HARDDISK, FwRev=0.12.1, SerialNo=QM1
Config={ Fixed }
RawCHS=16383/16/63, TrkSize=32256, SectSize=512, ECCbytes=4
BuffType=DualPortCache, BuffSize=256kB, MaxMultSect=16, MultSect=16
CurCHS=16383/16/63, CurSects=16514064, LBA=yes, LBAsects=73400320
IORDY=yes, tPIO={min:120,w/IORDY:120}, tDMA={min:120,rec:120}
PIO modes: pio0 pio1 pio2
DMA modes: sdma0 sdma1 sdma2 mdma0 mdma1 mdma2
UDMA modes: udma0 udma1 udma2 udma3 udma4 *udma5
AdvancedPM=no
Drive conforms to: ATA/ATAPI-5 published, ANSI NCITS 340-2000:
* signifies the current active mode
The bonnie numbers show for sequential output:
Good Machine Host: 76,857K/Sec
My Machine Host: 72,561K/Sec
Good Machine Centos 5.7 Guest: 66,266K/sec
My Machine Centos 5.7 Guest: 20,623K/sec
My machine Mandrake Guest: 1,365K/sec
Where should I look? I realize I do have two different front-ends to KVM, and
perhaps they are passing different parameters to it. I am also running the
KVM from Centos 6.2 on my machine, vs the other server is running on 5.7, but
I would have thought that newer is better. Also note that my hard drives
have a larger cache.
On a side note, I'm not thrilled with the Virtualizor's tech support, but the
product seems easy to use, once it actually works. Cloudmin seems to be
buggy, and not let you do things like change cd images on the fly, access the
console before the machine fully boots (!)... Any suggestions on other,
preferably open-source options? I'm a definite newbie to this virtualization
stuff.
Bob
___
CentOS mailing list
CentOS@centos.org
http://lists.centos.org/mailman/listinfo/centos
___
CentOS mailing list
CentOS@centos.org
http://lists.centos.org/mailman/listinfo/centos
Re: [CentOS] Really bad KVM disk performance
Bob Puff writes: /dev/hda: It should be /dev/vda.. it means you're not taking advantage of KVM's paravirtualisation capabilities. That probably explains the crappy performance. Run your disk and network with virtio and it should be much faster. In terms of management tools, check http://www.linux-kvm.org/page/Management_Tools hth -- Nux! www.nux.ro ___ CentOS mailing list CentOS@centos.org http://lists.centos.org/mailman/listinfo/centos
