Linux Virtualization

Kernel 2.6.23

Wed, 10 Oct 2007 09:19:00 EDT

Linux kernel 2.6.23 was released yesterday. Of note for virtualization is the addition of both Xen and lguest (see an introduction to lguest) support.

Xen support is based on the relatively new paravirt_ops interface, it is limited to 32-bit and guest only (no dom0 support).

Improvements have been made on the KVM front as well – guest SMP support and support for machines without PAE.

One non-virtualization feature that caught my eye is that the ext filesystem (ext4 specifically) finally has support for 32,000+ subdirectories – yes, I’ve seen braindead apps that need that. Of course it will be years before RHEL includes ext4 and most other distributions already have filesystems without this limitation.

Google Virtual Server Management

Mon, 3 Sep 2007 22:37:00 EDT

Recently appearing on Google Code is the Ganeti project.

Ganeti “is a virtual server management software tool built on top of Xen virtual machine monitor and other Open Source software”.

Ganeti supports 1-25 physical nodes, and HA using DRBD. It is interesting to see something based on DRBD, I had never considered it for production use but perhaps my reservations are unfounded.

openMosix project shutting down

Mon, 16 Jul 2007 14:37:00 EDT

Moshe Bar has announced that the openMosix project is ending because “The increasing power and availability of low cost multi-core processors is rapidly making single-system image (SSI) Clustering less of a factor in computing. The direction of computing is clear and key developers are moving into newer virtualization approaches and other projects”.

Moshe is behind the Qumranet startup that is developing Linux kernel virtual machine (KVM).

This loss of openMosix strikes me as a shame. I rather liked the idea of a single system image (extreme NUMA?), especially when combined with virtualization. Why not have a 64-way SMP box where groups of 4 to 8 CPUs are actually independent machines. CPUs could be added on the fly and parts of the “machine” could even be migrated to new hosts without impacting services.

I’m surprised that there isn’t demand / market for this but I guess there really isn’t, Virtual Iron used to offer single system image clustering (or should be call it aggregation) prior to their switch to the Xen hypervisor as well.

See announcement on openMosix mailing list and check out the openMosix website if you’re interested in more information.

RHEL5 Xen tutorial

Sun, 1 Jul 2007 11:31:00 EDT

Mickael Bailly has created a 5 part tutorial regarding his virtualization setup using RHEL5, Xen and GFS. It will be an interesting read for anyone who hasn’t used or studied Red Hat’s recent Xen release:

Enjoy.

Update: Unfortunately this presentation is offline. If anyone knows of an alternate location let me know, since I don’t host it I don’t control whether it’s available or not.

3 Leaf Systems introduces virtualization manager

Sun, 6 May 2007 15:37:00 EDT

3 Leaf Systems recently introduced their V-8000 Virtual I/O server. The product “replaces each compute node’s storage and network I/O with a single, high speed, redundant, and fault tolerant fabric, converting each compute node into a diskless and stateless commodity server with centrally managed bandwidth”.

Not a lot of information is publicly available (whitepapers are available on request), here’s a summary of what I’ve gathered:

all network and SAN I/O goes through central virtual I/O server
compute nodes and VIO server are interconnected with a 10Gb fabric (infiniband or ethernet)
consolidates 4-18 gigE network ports and 0-14 4Gb fibrechannel ports
VIO server runs RHEL4, 64-bit
compute nodes can run RHEL, SLES or Windows (limited versions, 32-bit and 64-bit)
in addition to the base operating systems (above) compute nodes can further virtualize things by running virtualization software including those based on VMware, Xen, and Microsoft technology

It is recommended to run the VIO servers in redundant pairs (no kidding!).

Sounds like an interesting solution. Check out their press releases here and the product datasheet.

RHEL4 Update 5 with Xen support

Wed, 2 May 2007 08:09:00 EDT

It looks like RHEL4 Update 5 is now out in some form. I didn’t see any official announcement yet but last night my internal yum repository got updated with a pile of RPMs which got me curious.

Anyway, whether Update 5 is official or not you can now grab an official Xen domU kernel for RHEL4 right from Red Hat.

Kernel errata package document is RHBA-2007-0304. You can grab either kernel-xenU-2.6.9-55.EL.i686.rpm or kernel-xenU-2.6.9-55.EL.x86_64.rpm.

Update: Red Hat’s release notes are here.

Tucows email service to run on Xen

Wed, 2 May 2007 06:54:00 EDT

Tucows Email Service is a large scale hosted email service and there is a new release in the works that is almost ready to be unleashed on the public (well at least the resellers).

In a recent blog post Tucows has let it slip that the new architecture is based on Sun X4600 servers running Xen virtual machines.

See Tucows Email Service: What’s Under the Hood? for the details – it is nice to hear about Xen getting some large production workloads.

Papers from recent Xen summit

Sun, 29 Apr 2007 12:02:00 EDT

Papers from the recently held Xen summit are now available online at http://xen.xensource.com/xensummit/xensummit_spring_2007.html.

With roughly 30 presentations online it’s a little tough to summarize, if you’re interested in Xen check it out.

FYI, it appears likely that support for paravirtualized Linux on Xen might finally make it into the next kernel (2.6.22), I won’t be holding my breath but things are finally sounding promising.

Linux kernel 2.6.21

Thu, 26 Apr 2007 23:51:00 EDT

Kernel 2.6.21 was released yesterday, a few of the more interesting highlights:

VMI now in default kernel (this is VMware’s layer on top of paravirt_ops), theoretically a fully paravirtualized Linux kernel on top of VMware might be do-able now
updated KVM code which includes some paravirt_ops support and live migration
clockevents and dynticks

The dynamic clock tick stuff is best covered over at LWN (see Clockevents and dyntick). This could have positive impact on power consumption and performance (particularly in virtualized environments). It will be interesting to see if certain virtual environments can now keep accurate time with a kernel such as this.

Lies, statistics and benchmarks

Thu, 8 Mar 2007 18:56:00 EST

VMware published a A Performance Comparison of Hypervisors at the end of January. Since the paper only discusses Windows I didn’t bother commenting – I don’t give two hoots about Windows.

Still the story is becoming amusing enough to warrant some study. Simon Crosby, Xensource CTO, recently published his thoughts the study:

All of XenSource’s commercial products match or beat ESX performance for Windows in all but a couple of benchmarks. This for a new product, and using the HVM feature set that has never been tuned. For Linux, we absolutely thrash ESX, which should come as no surprise. We’ll publish all of our results… just as soon as we get permission from VMware, that is.

Simon’s complete commentary is available in the Xensource blog here.

I would really have to agree with Simon, there are various commercial implementations of Xen (even if they aren’t allowed to be called that) and all will undoubtedly perform better than the older open source codebase that VMware did their comparison against.

Since VMware is comparing apples to oranges it reminds me a bit of the benchmarks I did a year ago with SuSE on Xen versus SuSE on ESX. Let’s just say embarrassing is not the word, joke might be the word but it sounds rather unprofessional – paravirtualization does help whether it’s just a paravirtualized network driver as in VMware’s windows benchmarks, or if it’s a completely paravirtualized kernel as in my SuSE on Xen tests.

Let’s hope that this evil policy of banning free and open benchmarking is lifted so that reasonable public discourse can take place without threat of lawsuit.