Submitted by “Moshe Bar, openMosix founder and project leader, has announced plans to end the openMosix Project effective March 1, 2008. The increasing power and availability of low cost multi-core processors is rapidly making single-system image 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.”
My opinion may not count, because I’ve never been able to properly setup a mosix cluster in the past (yes I’ve tried and failed).
However the latest available version information could say something:
Latest Release …
openMosix-kernel-2.4.26 December 9, 2004
No major release in more than two years, additionally no stable 2.6 kernel support (I don’t know if the support is in CVS).
Sorry, I always admired the project (even though I could not install it), yet the “end of life” decision seems to be appropriate.
I had an openmosix cluster of old junk PCs some years ago. It was kind of fun.
There is still the original Mosix, now called Mosix2, which supports 2.6. It’s proprietary, though.
http://en.wikipedia.org/wiki/Mosix
This is bad news. Multicore processors will never be able to replace putting hundreds of low power PC’s in my network into a single cluster.
Yeah. Too bad. Unfortunately, openmosix never quite worked that well; in the molecular interaction simulations that are run on the small cluster I administer, running several instances of the simulator in parallel has always been ridiculously faster than letting OM handle the migration. Even factoring in the HD speed (SCSI RAID) and the network (gigabit), the slowdown was surprisingly huge.
Not to mention only supporting the aging 2.4.x kernel and its sub-optimal CPU and I/O schedulers. 2.6.x never left the alpha/early-beta stages.
Well, what can we do. Consider he’s also the founder of xensource and qumranet. I guess he finally found where the real money is.
I can definitely say I saw that one coming. And I just can’t wait for DragonFly to become a viable alternative. There’s where the exciting single image cluster work is going on nowadays.
Clustering isn’t dead, it just needs to be reinvented for business computing as opposed to scientific computing. That means that clustering must become a part of the workload management solution starring virtualization. Clustering and virtualization are often seen as opposite solutions to opposite problems, but they can work together.
You’re probably familiar with the threading model debates from years past: 1:1, M:1, or M:N. They describe the mapping between user threads and kernel threads. But in the context of workload management, they can be used to describe the mapping between virtual and physical machines. In this sense, standalone is 1:1, virtualization is M:1, and clustering is 1:N.
If you combine virtualization and clustering, you get M:N, completely decoupling virtual machines from physical machines. In this scheme, the administrator can add virtual machines when additional functionality or software redundancy is needed, or she can add physical machines when additional capacity or hardware redundancy is needed.
It’s not about performance or parallelization, it’s about flexibility and manageability. It’s about pooling all of your computing resources together and then partitioning it up among workloads without worrying about the capacity of each box. It’s about a long-overdue marriage of centralized and distributed computing.
I think it will work, but not much of the OpenMosix codebase will be very useful anymore. And while DragonFly is a cool project, I see it as more of a successor to OpenMosix in the HPC market than a new clustering platform for enterprise IT. The new clustering functionality will grow out of partition migration and other HA-oriented virtualization features.
OpenMosix is about process migration. These days you can certainly relocate virtual machines, maybe losing extant network connections.
But it’s an http world, where connections are short lived anyway, and where client states are stored in server side sessions tied to client cookies.
Reclamation of resources is based on timeouts.
It is also about scaling. Well, for ecommerce, load balancing (round robin dns or network director) does the job.
The last habitat of OpenMosix-like systems is in clustered databases à la RAC, where transactional behavior and integrity are requirements.
Anyone have tested it these days? It looks nice, and got an update to kernel 2.6.20.
[1]http://www.kerrighed.org/
I remember a post on kerrighed on linuxfr which showed that it had very high performance (much better that the comparison).
But at the time, it wasn’t very scalable as any node crash would crash the system.
In the release note, they’re saying that they have disabled some features to concentrate on the reliability, a sound strategy, but I’m not sure what their current state is.
is there any valid alternatives?
Not sure if it is a Single System Image (SSI) solution like OpenMosix, but Beowulf has been around for ages and still is in widespread use:
http://www.beowulf.org
look at http://www.openssi.org
this is an alternative.
and it is actively developed. a new release should come in a few months as far as ai remember