Xen project Mailing List

Your "dd" test is fast because it writes data to kernel buffers, which flush asynchronously to disk.

Small tests are fast because the output file fits completely within disk block buffers. Large tests are slow because they exceed memory capacity, forcing buffers to flush to disk.

To see more realistic disk sequential throughput, run the same test with "oflag=direct".

From: xen-users-bounces@xxxxxxxxxxxxxxxxxxx [mailto:xen-users-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of Dot Yet
Sent: Friday, October 30, 2009 1:00 AM
To: xen-users@xxxxxxxxxxxxxxxxxxx
Subject: [Xen-users] I/O Performance concerns (Xen 3.3.2)

Hi everyone,

I need some help with regard to I/O performance on a xen based environment. Following is the setup.

Host Server Configuration:

CPU: 2 x 5430 Xeon CPUs

Motherboard: Tempest i5400PW S5397

RAM: 24GB RAM

Disks: 8 x 500 GB 7200 RPM disks - for VMs

SATA controller on 4 disks: Supermicro 8-Port SATA Card - (AOC-SAT2-MV8)

SATA controller on other 4 disks: onboard controller (Intel 6321ESB)

1 x 250 GB 7200 RPM disk - for OS

OS: OpenSolaris 2009.06 with Xen 3.3.2

The relevant zfs pool 'vmdisk' has been laid out as follows:

capacity operations bandwidth

pool used avail read write read write

---------- ----- ----- ----- ----- ----- -----

vmdisk 188G 1.18T 0 0 0 0

mirror 62.7G 401G 0 0 0 0

c8t2d0 - - 0 0 0 0

c9t0d0 - - 0 0 0 0

mirror 62.7G 401G 0 0 0 0

c8t3d0 - - 0 0 0 0

c9t1d0 - - 0 0 0 0

mirror 62.7G 401G 0 0 0 0

c8t4d0 - - 0 0 0 0

c9t4d0 - - 0 0 0 0

c8t5d0 0 464G 0 0 0 0

cache - - - - - -

c9t5d0 76.6G 389G 0 0 0 0

---------- ----- ----- ----- ----- ----- -----

Where c8t5d0 is a slog device and c9t5d0 is a cache device. The recordsize is default 128k and compression is set to OFF.

Xen dom0 has been limited to 2 GB RAM through following configuration in /rpool/boot/grub/menu.lst file:

kernel$ /boot/$ISADIR/xen.gz dom0_mem=2048M

domU has been configured as follows:

VCPUs - 4

RAM - 8GB

HDD - 100GB zvol on ZFS filesystem 'vmdisk'. The record size of the zvol is default of 8k and compression is set to OFF. Also used --nonsparse command line option while creating the VM through virt-install.

Paravirtualized

OS: CentOS release 5.4 (Final) X86_64

Kernel: 2.6.18-164.el5xen

Now, the problem. I am running the following dd command inside the domU:

[user1@db2db01 ~]$ dd if=/dev/zero of=j1 bs=8k count=100000

100000+0 records in

100000+0 records out

819200000 bytes (819 MB) copied, 2.57249 seconds, 318 MB/s

The above command returns under 2.5 seconds, however iostat on domU AND zpool iostat on dom0, both continue to show write IO activity for upto 30 to 40 seconds:

domU iostat:

avg-cpu: %user %nice %system %iowait %steal %idle

0.00 0.00 0.00 24.81 0.00 75.19

Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util

xvda 0.00 4989.00 0.00 1133.00 0.00 23.06 41.68 146.08 107.49 0.88 100.00

xvda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

xvda2 0.00 4989.00 0.00 1133.00 0.00 23.06 41.68 146.08 107.49 0.88 100.00

dm-0 0.00 0.00 0.00 6113.00 0.00 23.88 8.00 759.28 100.24 0.16 100.00

dm-1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

avg-cpu: %user %nice %system %iowait %steal %idle

0.00 0.00 0.00 24.94 0.00 75.06

Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util

xvda 0.00 4989.00 0.00 1153.00 0.00 23.91 42.47 146.32 146.37 0.87 100.40

xvda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

xvda2 0.00 4989.00 0.00 1153.00 0.00 23.91 42.47 146.32 146.37 0.87 100.40

dm-0 0.00 0.00 0.00 6143.00 0.00 24.00 8.00 751.75 143.43 0.16 100.40

dm-1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

dom0 zpool iostat:

capacity operations bandwidth

pool used avail read write read write

---------- ----- ----- ----- ----- ----- -----

vmdisk 188G 1.18T 287 2.57K 2.24M 19.7M

mirror 62.7G 401G 142 890 1.12M 6.65M

c8t2d0 - - 66 352 530K 6.49M

c9t0d0 - - 76 302 612K 6.65M

mirror 62.7G 401G 83 856 670K 6.39M

c8t3d0 - - 43 307 345K 6.40M

c9t1d0 - - 40 293 325K 6.40M

mirror 62.7G 401G 60 886 485K 6.68M

c8t4d0 - - 50 373 402K 6.68M

c9t4d0 - - 10 307 82.9K 6.68M

c8t5d0 0 464G 0 0 0 0

cache - - - - - -

c9t5d0 77.1G 389G 472 38 3.86M 3.50M

---------- ----- ----- ----- ----- ----- -----

capacity operations bandwidth

pool used avail read write read write

---------- ----- ----- ----- ----- ----- -----

vmdisk 188G 1.18T 75 3.52K 594K 27.1M

mirror 62.7G 401G 30 1.16K 239K 8.89M

c8t2d0 - - 10 464 86.6K 8.89M

c9t0d0 - - 19 350 209K 8.89M

mirror 62.7G 401G 0 1.18K 0 9.10M

c8t3d0 - - 0 510 0 9.10M

c9t1d0 - - 0 385 0 9.10M

mirror 62.7G 401G 45 1.18K 355K 9.11M

c8t4d0 - - 37 469 354K 9.11M

c9t4d0 - - 7 391 57.7K 9.11M

c8t5d0 0 464G 0 0 0 0

cache - - - - - -

c9t5d0 77.1G 389G 514 157 4.14M 17.4M

---------- ----- ----- ----- ----- ----- -----

Can you tell me why this happens? Is this behavior coming from Linux or Xen or ZFS? I do notice that iostat reports an iowait of 25%, but I don't know who is causing this bottleneck amongst them.

Instead of writing a 800 meg file, if I write an 8gb file, the performance is very poor (40 MB/s or so) and again, despite there is a long iowait after the dd command returns.

Any help would be really appreciated.

Kind regards,

dot.yet

RE: [Xen-users] I/O Performance concerns (Xen 3.3.2)