Linode.com Forum

[caker]

Pigs in Xen - The Linode Xen Beta Test

What is Xen, you ask? Xen is a virtualization technology that provides a few benefits over User Mode Linux, namely: performance, the ability to have SMP Linodes, and kernel modules.

How can I participate in the Xen beta test?

You may register for the Xen Public Beta by logging into the LPM and using the Xen Public Beta Registration page located off the "My Accounts" subtab (link is here: https://www.linode.com/members/xenbeta.cfm).

We will configure a migration to a Xen host which you can execute at your leisure. Migrating moves your Linode's disk images, config profiles, and IP addresses to the new host. The migration is seamless.

What would you like us to test during the beta?

Our Xen implementation has been in development off and on for a few years now (!), so it's well tested and feature complete. While I don't anticipate any bugs, the point of the beta test is to prove that.

Our focus will be mainly on performance and resource fairness. Xen has some sophisticated tools to weigh the scheduling of CPU time, and we'll be using the ionice feature of the CFQ disk scheduler instead of the token-limiter patch that I wrote for UML. All that needs to be tested. Mainly, I want you to continue using your Linode as you did -- under normal, production loads.

When will the Xen Public Beta end, and what happens then?

The beta test is open ended. It won't be until we're confident in all aspects of Life Under Xen (functionality, performance, reliability, etc) that we'll end the beta.

During the beta test, Linodes of different plans will be on the same host. Once the beta test ends, some of you will need to migrate off onto production Xen boxes with other Xen Linodes of the same plan. That will be another one-click migration.

I found a bug, where should I report it?

Post a new thread in this forum. That way, we can verify and track any bugs that are found.

Is there a Linode-Xenbeta IRC channel?

Yup. #linode-xenbeta on OFTC.

http://www.linode.com/irc/

Wait a second. I thought Xen could use NPTL!

It can, but with a performance hit. Feel free to turn off the "disable /lib/tls" helper and move /lib/tls.disabled/ back to /lib/tls. Debian has a "libc6-xen" package you can install, or build libc with "-mno-tls-direct-seg-refs".

Will there be 2.4 and 2.6 Xen kernels?

Xen 3.0 only supports Linux-2.6.

What doesn't work yet?

All the features of the Linode Manager work as they did under UML.

-----

If there are other questions, please let me know and I'll add them to this announcement.

Thanks!
-Chris

[zunzun]

caker wrote:
...namely: performance, the ability to have SMP Linodes,
-Chris

I signed up since my Linode is such a CPU hog. I'll be all over SMP.

James

[zunzun]

caker wrote:
...namely: performance, the ability to have SMP Linodes,
-Chris

Well, that was easy.

Performance with apt-get is noticeably better. Floating point performance is quite good, the Ubuntu atlas3-sse2-dev (Automatically Tuned Linear Algebra Subroutines with CPU SSE2 extensions) installed natively with apt-get and runs well giving excellent linear algebra performance. Python pickling/depickling of complex objects is speedy.

I do not see multiple CPU's (with top) for SMP testing, drool drrrroooool slobber...

James

[caker]

zunzun wrote: I do not see multiple CPU's (with top) for SMP testing, drool drrrroooool slobber...
It's coming. I'm trying to get a kernel newer than what Xen comes with (2.6.18 ) to work -- but it may have to wait until tomorrow. I'll let you guys know once it's working...

-Chris

[zunzun]

I use Linux 'nice'-ness to control multi-process CPU allocation, and interestingly this works considerably better on the Xen beta server. That is, the kernel scheduler seems to pay very close attention to the 'nice' level now - I had not seen this fine a granularity of CPU allocation previously. This is excellent, and will be much better for users of my web site.

James

[OverlordQ]

hot.

[caker]

OK, we're good to go with the paravirt_ops guest kernel, which is currently 2.6.23.8-domU-linode2.

Enjoy,
-Chris

[zunzun]

caker wrote: OK, we're good to go with the paravirt_ops guest kernel, which is currently 2.6.23.8-domU-linode2.


Well, I see 3 processes at 99%CPU each - ahhhh, so lovely! Seems to work perfectly.

James

[zunzun]

6 processes at 66% each (4/6) looks correct for 4 cores, so load balancing looks good from here.

James

[zunzun]

The 'nice'-ness behaviour seems inverted - process with high nice numbers are being given more CPU that those with low nice numbers.

James

[zunzun]

zunzun wrote: The 'nice'-ness behaviour seems inverted - process with high nice numbers are being given more CPU that those with low nice numbers.


Well, it's working properly now - must have been my test. The box gets 4 stars out of 4 from me on both performance and resource allocation.

James

[egatenby]

Is any space still available in Fremont for the Xen beta test?

[caker]

egatenby wrote: Is any space still available in Fremont for the Xen beta test?
There is, but we've got two outstanding bugs in Xen we're waiting to get resolved before setting up the pending Xen beta requests, since one of the bugs is critical. Otherwise, the beta test has been going well.

Feel free to sign up for the beta test now and we'll inform you once things are cleared up.

-Chris

[zunzun]

caker wrote:
...the beta test has been going well.


I'd mark that as 'very well', personally. Only one minor hiccup.

James

[jcr]

Quote: we'll be using the ionice feature of the CFQ disk scheduler instead of the token-limiter patch that I wrote for UML
Does that mean there are no limits to IO? Assuming, there are some limitations, is there a way to monitor the IO situation? Something like the /proc/io_status file?