Refute of Mr. Walker's White Paper on openMosix vs. OpenSSI

With regard to “A Comparison of OpenSSI and OpenMosix” by Bruce J. Walker, we wish to refute the author’s claims and present the facts. Mr. Walker’s document was presented as a white paper comparison of the two clustering products; however, the author lacks an adequate command of openMosix to make such a comparison.

First, let us state clearly that we are openMosix, not OpenMosix.

Mr. Walker states that openMosix does not present a Single System Image. This is not true. All of the resources of the cluster are available to any node of an openMosix cluster. A process’ home node is simply the node where a process originally started. There is no single Home node for an openMosix cluster.

openMosix is designed for high performance computing not high availability. The objective is total throughput (speed) for the cluster. Problems presented to openMosix are usually long running calculations that can be rerun without much to do.

For processes that are very long running, there is a very capable checkpoint and restart in openMosix, called CHPOX. This checkpoint/restart is so well done that the processes could actually be moved to another similar cluster and restarted. In other words we have the full migration that Mr. Walker talks about, too, we simply don't see it as as a big feature.

Shared memory is also available to openMosix through another openMosix. For many HPC clusters, shared memory is not an important feature, but we have it nonetheless and continually improve the performance of it which we feel is superior to others we have seen.

Mr. Walker presents openMosix’ migration of memory pages as a disadvantage over migration of entire memory space for a process. For large programs, the time to migrate an entire program even on Gigabit Ethernet or InfiniBand can be disturbingly long. One could conceivably use the entire cluster bandwidth before all of the nodes had programs to run. Full migration cannot be used to load balance large apps, which are increasingly in use today.

Mr. Walker implies that openMosix is lacking in membership API and node management, but openMosix has a very strong API with membership of nodes in the cluster done automatically. openMosix Contributing Developers provide alternate monitoring and management products for use with openMosix; however, none is actually required if a cluster owner wishes to not use them.

openMosix is a core product much more like an operating system than an application. That allows openMosix to appear much different from cluster to cluster. For example, several developers have created instant cluster products that boot a CD/DVD into a running openMosix cluster complete with GUI monitoring and management software, shared memory, security, and many applications that can take advantage of the cluster. Still, the same openMosix kernel patch can be installed for a HPC cluster that appears to be an entirely different product. This is by design and is not a short coming of openMosix. We are a platform and we have dozens of contributors that build on top of it. There are even some companies that base some of their software on openMosix (see Qlusters.com).

Mr. Walker points out that MFS (actually it is oMFS) is rather limited compared to some of the better cluster file systems available today. He has not bothered to notice that oMFS has been dropped from openMosix. openMosix users are currently using a wide variety of file systems. And, the openMosix Project is moving to provide easier integration with several of the better file systems.

Minor refutes are furthermore:

We do have an mfork feature contrary to Mr. Walker's assertion and we do have a cluster membership function which doesn't need any user intervention. openMosix is completely config file free. It auto configures itself.

We at openMosix feel we have a valid platform that finds traction in the market. We have over 30,000 installed clusters world wide, with users such as NASA, U.S. Navy, U.S. Air Force, NSA, MIT, Yale, Harvard, Morgan Stanley, Deutsche Bank, Bear Stearns, IBM, Boeing, U.N. Atomic Agency, Ferrari, Israeli defense forces. And, there are many, many more. Just compare Google results: 16,900 OpenSSI vs. 179,000 openMosix.

We have such a consistently rising user base because we target a single market, HPC. We do not try to put in bells and whistles where not needed. Also, we add the features ourselves to have harmonic integration and do not integrate with other technologies and then claim they are part of our technology.