Posted by: Anonymous Coward
on August 27, 2003 09:44 PM
This article presents an extremely limited view of Linux clusters. It ignores anything about networking among them, and ability for machines to work together to solve problems.
The earliest and simplest clusters used the approach described here of parceling out individual jobs to individual nodes for each one to solve independent of the others. But since NASA invented Beowulf about ten years ago, tools such as MPI and PVM have been used to run parallel jobs across an entire Linux cluster (even longer than that if one includes non-Linux clusters).
And Mosix allows transparent use of threaded code on clusters: if multiple threads or processes are running on one machine, it migrates them to others, and uses the network to create the illusion of shared memory among all threads. In other words, anything that will run on a "supercomputer" will run on a Mosix cluster, with the only differentiating factor being the relative speeds of communication between nodes. And though a handful of applications require the extremely fast communication of a "supercomputer" like the SGI Altix (e.g. density functional theory calculations for electronic structure), a great many can run at full processor utilization even with limited network bandwidth.
This ability to run parallel jobs has been the focus of Beowulf supercomputing since the beginning, and it's sad to see it being omitted from a piece like this in 2003. The author would do well to read up on the subject before writing about it publicly; I recommend the June 2003 issue of Linux Magazine, which has a variety of historical and other information about clusters.
Learn about Linux
Download Linux
Get Linux help
Ever heard of MPI or PVM or Mosix or ...
Posted by: Anonymous Coward on August 27, 2003 09:44 PMThe earliest and simplest clusters used the approach described here of parceling out individual jobs to individual nodes for each one to solve independent of the others. But since NASA invented Beowulf about ten years ago, tools such as MPI and PVM have been used to run parallel jobs across an entire Linux cluster (even longer than that if one includes non-Linux clusters).
And Mosix allows transparent use of threaded code on clusters: if multiple threads or processes are running on one machine, it migrates them to others, and uses the network to create the illusion of shared memory among all threads. In other words, anything that will run on a "supercomputer" will run on a Mosix cluster, with the only differentiating factor being the relative speeds of communication between nodes. And though a handful of applications require the extremely fast communication of a "supercomputer" like the SGI Altix (e.g. density functional theory calculations for electronic structure), a great many can run at full processor utilization even with limited network bandwidth.
This ability to run parallel jobs has been the focus of Beowulf supercomputing since the beginning, and it's sad to see it being omitted from a piece like this in 2003. The author would do well to read up on the subject before writing about it publicly; I recommend the June 2003 issue of Linux Magazine, which has a variety of historical and other information about clusters.
#