Optimize KVM performance

goexodus

Was trying find ways to optimize KVM installs Centos 6.3 or Ubuntu 12.04 and could not find any good information except:

1) modify /boot/grub/menu.lst or /boot/grub/grub.cfg to allow VPS IO scheduler to noop

2) use Virtio Drivers

3) fix Swap Partition if its 0
free -m
swapoff /dev/vda2
mkswap /dev/vda2
swapon /dev/vda2

4) improve I/O performance
echo 0 > /sys/block/vda/queue/rotational
echo 0 > /sys/block/vda/queue/rq_affinity
echo noop > /sys/block/vda/queue/scheduler
echo "echo 0 > /sys/block/vda/queue/rotational" >> /etc/rc.local
echo "echo 0 > /sys/block/vda/queue/rq_affinity" >> /etc/rc.local
echo "echo noop > /sys/block/vda/queue/scheduler" >> /etc/rc.local
echo 'vm.swappiness=5' >> /etc/sysctl.conf
echo 'vm.vfs_cache_pressure=50' >> /etc/sysctl.conf
sysctl -p

5) mount partition with noatime.
vi /etc/fstab

any other links, tips or ideas?

Ramnode seems to have a well done guide
https://clientarea.ramnode.com/knowledgebase.php?action=displayarticle&id=44

Bogdacutuu

there's not much you can do to optimize KVM performance, what issues do you have right now?

concerto49

Move to a faster node

goexodus

there's not much you can do to optimize KVM performance, what issues do you have right now?

virtio driver instead of the legacy IDE made a noticable difference at least on Centos.

Bogdacutuu

@goexodus said: virtio driver instead of the legacy IDE made a noticable difference at least on Centos.

yeah, I meant except the methods you already described

Keith

Use ext4 instead of ext3 when installing.
The kernel could also be recompiled, my choices are
Not compiling for size
Preemption model is No Forced Preemption (Server)
Processor family - depends on cpu used
Reduce Maximum number of CPUs

yomero

None of that instructions is oriented to SSDs except this

echo 0 > /sys/block/vda/queue/rotational

The nobarrier option seems to be suggested because they use BBU raid controllers.
The noatime isn't really recommended, apparently "relatime" is a better choice.
The noop scheduler is the recommended for KVM too.

@Keith said: Use ext4 instead of ext3 when installing.

Why this?

fly

noop is good fo kvm

Ash_Hawkridge

@fly said: noop is good fo kvm

+1

Nick_A

If you're having any issues on a RamNode VPS, don't hesitate to open a ticket.

Otherwise, carry on! Perhaps I'll learn a thing or two in this thread as well.

George_Fusioned

Adding the following lines to your /etc/sysctl.conf file will increase network throughput

net.core.rmem_max=16777216
net.core.wmem_max=16777216
net.ipv4.tcp_rmem=4096 87380 16777216
net.ipv4.tcp_wmem=4096 65536 16777216

You can also try the following /etc/fstab options on your ext4 mount

noatime,data=writeback,barrier=0,nobh

yomero

I don't like the noatime one =/

Also, I am afraid of the cache one.

goexodus

But save all settings in case you need rollback ...

Nick_A

I've found that double writeback is not helpful. Just let the hardware handle writeback in my opinion.

yomero

@Nick_A said: I've found that double writeback is not helpful. Just let the hardware handle writeback in my opinion.

So, at the end, the "perfect" settings depend of the provider :P

George_Fusioned

@Nick_A said: I've found that double writeback is not helpful. Just let the hardware handle writeback in my opinion.

I believe this isn't a provider specific topic, so there's always a chance someone's provider not using writeback.

Nick_A

Sure, sorry.

Anyway, have you guys seen a KVM guest become very slow after applying tweaks like the ones above?

goexodus

Does higher count of threads or CPUs have any predictable effect or its worse as in the case of VMWare?

yomero

@goexodus said: as in the case of VMWare?

:O What you mean?

goexodus

VMWare does not behave well when you assign multiple vCores at multtiple VMs. You actually pay a penalty in that case. VMs dispatches vcores using a time slicing algorithm (fair approach), not an interrupt driven algorithm as is typically used to dispatch processes in a regular operating system. What this means in practice is that if you assign say 2 vcores to a guest, then 2 vcores must be AVAILABE AT SAME TIME before that guest can run at all.

I was just wondering how to offer increased benchmarks in certain customers they need it and willing to pay using KVM VM platform?

yomero

@goexodus said: AVAILABE AT SAME TIME before that guest can run at all.

Wow, that sounds a little bit dumb

goexodus

Wow, that sounds a little bit dumb

No its not actually, Quite the opposite. Its a complicated subjects and statistical in nature ...

fileMEDIA

@goexodus You get more performance with more vCores. Take a look at serverbears benchmarks on the unix bench results, if you have more vCores you get more points. It works very well with linux kvm.

You can pin each vCore to a specific core if you want to get the the best results.

yomero

@fileMEDIA said: You get more performance with more vCores.

Which sounds obvious lol

@goexodus said: Its a complicated subjects and statistical in nature

I guess so. VMware stuff has years and years of development

goexodus

I guess so. VMware stuff has years and years of development

But if you check their licencing you will get a heart attach ....