Hi, I just run newer kernels because I hoped it would improve the IO performance, which apparently is not the case. Which kernel versions are you running? Then I can try those, to see if that solves the problem. Best, -- Jaap Van: Xen-users <xen-users-bounces@xxxxxxxxxxxxxxxxxxxx> Namens Roalt Zijlstra | webpower
Verzonden: Tuesday, 18 December 2018 16:12
Aan: Håkon Alstadheim <hakon@xxxxxxxxxxxxxxxxxx>
CC: xen-users@xxxxxxxxxxxxxxxxxxxx
Onderwerp: Re: [Xen-users] Disk performance on guest incredibly low. You are definitely running newer kernels then I do.. But I tested that with a setting in the grub.cfg I can enable the multi-queue block mode by adding 'scsi_mod.use_blk_mq=1'. I would think that if you use 'scsi_mod.use_blk_mq=0' you can disable the multi-queue block mode. In general the multi-queue block modes should be better, but it is worth a try in combination with Xen to test the single queue block mode. | Barcelona | Barneveld | Beijing | Chengdu | Guangzhou
Hamburg | Shanghai | Shenzhen | Stockholm | |
|
Den 17.12.2018 10:32, skrev Roalt Zijlstra | webpower:
> Hey Håkon,
>
> One little tip on disk IO in Virtual hosts, is changing the FS
> scheduler from cfq to noop. In our Xen PV configs we add this to the
> xen.cfg files at the end:
>
> extra="clocksource=tsc elevator=noop"
>
> Especially the "elevator=noop" parameter is forcing all FS-es to the
> noop scheduler. In my experience that gave our Xenservers a pretty
> nice boost.
> Red Hat does recoomend this for all Virtual servers (see the link
> below). To test this you don't need to reboot at all.
>
> For example from the link below there is a code snippet about getting
> and setting the scheduler for /dev/hda. (Replace hda with sdb or any
> other block device)
>
> # cat /sys/block/hda/queue/scheduler
> noop anticipatory deadline [cfq]
>
> # echo 'noop' > /sys/block/hda/queue/scheduler
> # cat /sys/block/hda/queue/scheduler
> [noop] anticipatory deadline cfq
>
> More info: https://access.redhat.com/solutions/5427
Yes, found that resource some time ago. Tried again now, no
break-through. I've got 'none' available as scheduler, rather than noop,
but they should be equal. No matter what I do (in dom0 and/or domu) I
get at least 10 x higher speed in the dom0. :-/ .
Example:
###
###In dom0, md-raid on drives f j g h i k. echo-and-cat does the obvious.
### (I usually run with mq-deadline as scheduler. seems to give
marginally better performance)
# for f in f j g h i k ; do echo-and-cat
/sys/block/sd${f}/queue/scheduler;done
/sys/block/sdf/queue/scheduler
[none] mq-deadline kyber bfq
/sys/block/sdj/queue/scheduler
[none] mq-deadline kyber bfq
/sys/block/sdg/queue/scheduler
[none] mq-deadline kyber bfq
/sys/block/sdh/queue/scheduler
[none] mq-deadline kyber bfq
/sys/block/sdi/queue/scheduler
[none] mq-deadline kyber bfq
/sys/block/sdk/queue/scheduler
[none] mq-deadline kyber bfq
# mount /dev/disk/by-label/SAS-STEAM /mnt/tull
# cd /mnt/tull/tmp
# df -hT ./
Filesystem Type Size Used Avail Use% Mounted on
/dev/md2p8 ext4 196G 164G 23G 88% /mnt/tull
# dd if=/dev/zero of=a_file bs=1M count=1024 conv=fsync oflag=direct
1024+0 records in
1024+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 4.56845 s, 235 MB/s
0:root@gentoo tmp # dd if=/dev/zero of=a_file bs=1M count=1024
conv=fsync oflag=direct
1024+0 records in
1024+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 7.52557 s, 143 MB/s
###
### Booting domu with root and /tmp on SAS-STEAM, ssh into domu
###
# cd /tmp
# df -hT ./
Filsystem Type Størrelse Brukt Tilgj. Bruk% Montert på
/dev/xvdb ext4 196G 163G 24G 88% /
# cat /sys/block/xvdb/
alignment_offset device/ holders/ power/
ro subsystem/
bdi/ discard_alignment inflight queue/
size trace/
capability ext_range integrity/ range
slaves/ uevent
dev hidden mq/ removable stat
# cat /sys/block/xvdb/queue/scheduler
[none] mq-deadline
# cd /tmp
# df -hT ./
Filsystem Type Størrelse Brukt Tilgj. Bruk% Montert på
/dev/xvdb ext4 196G 163G 24G 88% /
# cat /sys/block/xvdb/
alignment_offset device/ holders/ power/
ro subsystem/
bdi/ discard_alignment inflight queue/
size trace/
capability ext_range integrity/ range
slaves/ uevent
dev hidden mq/ removable stat
# cat /sys/block/xvdb/queue/scheduler
[none] mq-deadline
# dd if=/dev/zero of=a_file bs=1M count=1024 conv=fsync oflag=direct
1024+0 oppføringer inn
1024+0 oppføringer ut
1073741824 byte (1,1 GB, 1,0 GiB) kopiert, 64,2402 s, 16,7 MB/s
# dd if=/dev/zero of=a_file bs=1M count=1024 conv=fsync oflag=direct
1024+0 oppføringer inn
1024+0 oppføringer ut
1073741824 byte (1,1 GB, 1,0 GiB) kopiert, 59,7517 s, 18,0 MB/s
#
|
