Web Hosting Talk : Web Hosting Main Forums : VPS Hosting : Paravirtualization vs Virtual Containers

[Yash-JH]

This is an interesting study conducted by University of Michigan and HP comparing XEN (paravirtualization) and OpenVZ (virtual containers).
XEN, MS Virtual Server and VMWare all use paravirtualization, while OpenVZ and Virtuozzo use Virtual Containers.

The study reports that response times can increase by 400 to 600% for XEN, even with the host running only a few nodes. This is largely because of the overhead XEN presents. OpenVZ had a minimal response time increase of around 100% for the same workload. The study also noted that XEN used twice the amount of CPU for the same workload compared to OpenVZ

What the article says is that L2 cache misses (both memory and instruction) signficantly increased in the XEN study. For non Computer Science people, that basically means the processor caching system was far less effective for XEN, compared to OpenVZ (and single OS systems). Caching is important because the time it takes to read instructions or data from RAM is significantly slower than reading instructions from the cache. Cache misses are similar to page-faults (in OS memory architecture).

To a consumer what this means is that Virtual Containers (OpenVZ, Virtuozzo) may offer much better performance than any paravirtualization based solution as long as the host isn't overloading their VPS system.

While paravirtualization offers higher fault-tolerance, in most well designed virtual container based VPS systems, this is minimized and the performance overhead may not be worth it.

Comments?
http://www.hpl.hp.com/techreports/2007/HPL-2007-59.pdf

[devonblzx]

Yes, I had seen this with my own testing that OpenVZ put out better performance which is why I had opted the VZ route a long while ago.

Maybe we can hear from all the hosts who have claimed Xen performs better on this one ;-).

[vibrokatana]

I have a xen VPS and am entirely happy with it. I went through 2 Virtuozzo hosts and ended up with wacky issues like network pauses, incredibly slow ssh and just general slowness. My guess is that hosts are overselling these machines by throwing in loads of ram in each box, and virtually none of them state the share ratio.

[Yash-JH]

Most certainly, these issues you decribe vibrokantana would be from a bad VPS host (too many nodes per host). Unfortunately, many low-cost VPS hosts will exploit the low overhead by putting many VPS accounts per server.. hence degrading performance.

A good VPS host will not do this. But conceptually, if done right, Virtuozzo would be far superior than any paravirtualization solution

[vibrokatana]

Originally Posted by Yash-JH

Most certainly, these issues you decribe vibrokantana would be from a bad VPS host (too many nodes per host). Unfortunately, many low-cost VPS hosts will exploit the low overhead by putting many VPS accounts per server.. hence degrading performance.

A good VPS host will not do this. But conceptually, if done right, Virtuozzo would be far superior than any paravirtualization solution

The current provider I am with is 1:11 share ratio, 256MB of ram and 15GB of disk space for 25 dollars a month. Of course that is unmanaged and no control panels thrown in (I love managing servers tho so it all works out good).

[Yash-JH]

That would be a good ratio for Virtuozzo. However for XEN, based on that study you could be experiencing performance degradation compared to a Virtuozzo host with the same ratio

[vibrokatana]

Originally Posted by Yash-JH

That would be a good ratio for Virtuozzo. However for XEN, based on that study you could be experiencing performance degradation compared to a Virtuozzo host with the same ratio

Of course you will have fun actually finding a virtuozzo host with the same ratio/performance that wont cost you an arm and leg.

[utropicmedia-karl]

We use VMWare and couldn't be happier - our clients say the same. I will say you can't get great performance out-of-the-box, but a few hundred hours of tweaking, custom kernels and learning how things work and we have native speeds.



Kind Regards,

[RaghavT]

This is pretty good study....

Thanks

[ServerDave]

The linked PDF and study are from April 2007. Not sure how applicable the data is to current systems. It would be interesting to see results on more recent software versions.

[ctaborda]

This threads become endless.

[ctaborda]

OpenVZ folks will say their approach allows for better resource usage and density. We feel paravirtualization provides good resource usage and superb isolation. Translation: your neighbor can’t bring the whole box down.

Xen also requires fixed memory and disk definitions, OpenVZ allows for burstable memory usage. That is the biggest difference you will see, burstable memory rates on Virtuozzo offerings, whereas Xen has hard, fixed caps. Burstable memory is great if you have control over all of the virtual servers (everyone is friendly), but when you have a diverse environment, we prefer hard memory caps (you get what you pay for).

Also, Xen gives users much more control over how his/her VPS is going to behave. Kernel swapiness? Customised iptables module? A different file system? Also a kernel level local exploit will only bring down your node, not everyone else's.

source: slice forums.

I couldn't have said it better.

[Yash-JH]

Originally Posted by vibrokatana

Of course you will have fun actually finding a virtuozzo host with the same ratio/performance that wont cost you an arm and leg.

I am sure you could, at the same price range.

The linked PDF and study are from April 2007. Not sure how applicable the data is to current systems. It would be interesting to see results on more recent software versions.

This study should hold for most intel x86 systems most Windows VPS providers are using. Infact, the processors/hardware most VPS hosts are running XEN and VMWare on are terrible and unsuited for such virtualization

We use VMWare and couldn't be happier - our clients say the same. I will say you can't get great performance out-of-the-box, but a few hundred hours of tweaking, custom kernels and learning how things work and we have native speeds.

The basic issue is that processor architecture is not designed to deal with the caching requirements of a system running paravirtualization. The processor's caching designs are for a single OS setup

When multiple OS's (and multiple applications within that OS) are continually switching back and forth, fighting for the CPU, you are bound to have massive L2 cache misses... resulting a severe response time degradation

[Yash-JH]

Originally Posted by ctaborda

Xen also requires fixed memory and disk definitions, OpenVZ allows for burstable memory usage. That is the biggest difference you will see, burstable memory rates on Virtuozzo offerings, whereas Xen has hard, fixed caps. Burstable memory is great if you have control over all of the virtual servers (everyone is friendly), but when you have a diverse environment, we prefer hard memory caps (you get what you pay for).

Virtuozzo (based on OpenVZ) offers plenty of software devices to maintain virtual containers, hence offering high levels of isolation. paravirtualization is the highest form of isolation, but virtual containers in most setups will rival it.

But my major point is with such massively degraded response times of around 400%, is the additional isolation worth the performance factor?

[ctaborda]

Most, if not all of our customers from an OpenVZ or Virtuozzo environment. All we hear is how great their Xen VPS runs vs their old vps.

In our VPS their application, might run slow if they use up too much ram, but in OpenVZ it will just crash.

Then, as a provider you start getting pissed at a VPS customer using too many resources, etc.

Xen makes it better to sleep at night.