Clustering
This is a good site http://www.beowulf.org/overview/howto.html
Follow the link towards the bottom to read on - his book goes to some lengths to describe the limitations and reasons why its not really worth the effort, unless you have a program designed and optimised for a parallel processing environment.
Some programs are getting better at making use of parallel processing, but generally this only extends as far as running as multiple processes rather than a single one, with processes being assigned to cores in a similar way to that described by Samuria in reference to the simplistic way that TS Sessions are load balanced amongst a cluster.
Basically its not worth the effort for the majority of programs.
It won't work the way you hope - it will of course allow you to run something like VMWare and then have say and XP VM sat in there, but for the most part you won't see much performance increase (although you will see a huge leap in power consumption, as well as having quite a bit of warmth to dispose of!)
Follow the link towards the bottom to read on - his book goes to some lengths to describe the limitations and reasons why its not really worth the effort, unless you have a program designed and optimised for a parallel processing environment.
Some programs are getting better at making use of parallel processing, but generally this only extends as far as running as multiple processes rather than a single one, with processes being assigned to cores in a similar way to that described by Samuria in reference to the simplistic way that TS Sessions are load balanced amongst a cluster.
Basically its not worth the effort for the majority of programs.
It won't work the way you hope - it will of course allow you to run something like VMWare and then have say and XP VM sat in there, but for the most part you won't see much performance increase (although you will see a huge leap in power consumption, as well as having quite a bit of warmth to dispose of!)
VMWare clusters tend to be for HA (High Availability) rather than for performance improvements.
It would be cheaper to buy a faster computer than it would to create a cluster to provide high-performance computing. (In terms of VMWare ERX Licenses)
Unless you buy the Enterprise version, you will more or less be stuck with each VM running on a single node in your cluster I think.
There are huge wins to be had in parallel processing (think about how great Google is at returning inordinate amounts of data in next to no time) To take advantage, you need to be doing LOTS of processing, and you need to be running apps designed ground up for a parallel environment.
Why not give it a go on a single node beowulf 'cluster' and benchmark, then expand your cluster to two nodes and benchmark again. Then try three nodes, and check again what improvment you have (if any).
I suspect performance will decrease as you add nodes, but would be happy to be proved wrong!
Go on, do the experiment, for the 'greater good'!
It would be cheaper to buy a faster computer than it would to create a cluster to provide high-performance computing. (In terms of VMWare ERX Licenses)
Unless you buy the Enterprise version, you will more or less be stuck with each VM running on a single node in your cluster I think.
There are huge wins to be had in parallel processing (think about how great Google is at returning inordinate amounts of data in next to no time) To take advantage, you need to be doing LOTS of processing, and you need to be running apps designed ground up for a parallel environment.
Why not give it a go on a single node beowulf 'cluster' and benchmark, then expand your cluster to two nodes and benchmark again. Then try three nodes, and check again what improvment you have (if any).
I suspect performance will decrease as you add nodes, but would be happy to be proved wrong!
Go on, do the experiment, for the 'greater good'!
The basic premise of a Beowulf cluster is HPC - High Performance Computing.
One very good resource is this.
Standard architecture of a Beowulf cluster is to have one "controller" node which parcels out processing tasks for the individual nodes that it controls. Depending on the type of application, you would have maybe one controller node per 16 to 256 'work' nodes. Each of these "compute farms" can then be controlled by a master controller node, which may control from 4-16 "farms". Each of these farm controllers could also be controlled by a master controller, which might look after the work assignments of 4-16 'farms' and so on. One very good illustration of a Beowulf type application is a CGI Render farm (incidentally, my new job as of 5th January is in charge of looking after the CGI render farm at Pinewood Studios!)
The image on the left shows one Beowulf "farm"
Unfortunately, running any sort of commodity software on a Beowulf cluster is pretty much a physical impossibility. Software does have to be specifically written using a technique based upon "message passing" which allows long and complicated calculations to be broken out into individual segments that can be pushed out to an individual node to complete.
So, the only real reason to build a Beowulf cluster (apart from the geek factor, which I totally understand ! ) is if you have access to programmers who can write software that is targetted for this sort of environment.
However
There's always a 'but'.
There is a software package called MOSIX which will allow you to use a cluster of computers as a single compute farm. Essentially, you set up a bunch of computers running Linux, and install MOSIX on them. Then, you can start any process on any one of the computers, and MOSIX will automatically find the best "resource" for that program to run on. MOSIX does work at the Process level, so it might allow you to spread the load of running a VM over multiple machines, but I really couldn't comment on how well this would work. There are a few caveats, such as you need to have shared storage (like a NAS box, or another machine running OpenFiler for example). The MOSIX site makes for some very interesting reading though!
Hope that helps!
B.
One very good resource is this.
Standard architecture of a Beowulf cluster is to have one "controller" node which parcels out processing tasks for the individual nodes that it controls. Depending on the type of application, you would have maybe one controller node per 16 to 256 'work' nodes. Each of these "compute farms" can then be controlled by a master controller node, which may control from 4-16 "farms". Each of these farm controllers could also be controlled by a master controller, which might look after the work assignments of 4-16 'farms' and so on. One very good illustration of a Beowulf type application is a CGI Render farm (incidentally, my new job as of 5th January is in charge of looking after the CGI render farm at Pinewood Studios!)
The image on the left shows one Beowulf "farm"
Unfortunately, running any sort of commodity software on a Beowulf cluster is pretty much a physical impossibility. Software does have to be specifically written using a technique based upon "message passing" which allows long and complicated calculations to be broken out into individual segments that can be pushed out to an individual node to complete.
So, the only real reason to build a Beowulf cluster (apart from the geek factor, which I totally understand ! ) is if you have access to programmers who can write software that is targetted for this sort of environment.
However
There's always a 'but'.
There is a software package called MOSIX which will allow you to use a cluster of computers as a single compute farm. Essentially, you set up a bunch of computers running Linux, and install MOSIX on them. Then, you can start any process on any one of the computers, and MOSIX will automatically find the best "resource" for that program to run on. MOSIX does work at the Process level, so it might allow you to spread the load of running a VM over multiple machines, but I really couldn't comment on how well this would work. There are a few caveats, such as you need to have shared storage (like a NAS box, or another machine running OpenFiler for example). The MOSIX site makes for some very interesting reading though!
Hope that helps!
B.
