The price of “legacy” systems is dropping at a fast rate. Quad PIII servers can be purchased on Ebay for less than $800. Stacks of old SGI systems (Indy and Indigo) systems are being sold for rock bottom prices. But until one single unified API is created to unite all these systems into one single cluster for use by the common computer user, they will remain useless to most people. This is where HPVM comes in. HPVM exists to bring the cluster concept out of the Universities and research facilities and make them useful for everything from video editing, to database servers.
Heterogeneous Parallel Virtual Machine
About The Author
Eugenia Loli
Ex-programmer, ex-editor in chief at OSNews.com, now a visual artist/filmmaker.
Follow me on Twitter @EugeniaLoli
Great idea! Looks like its in the early stages though. Hopefully they can keep the API simple so that useful software can be deveolped for it.
Isn’t this exactly what the Parallel systems built on Java already achieve?
The hidden agenda of HPVM is to create the simplest interface possible and still retain some sort of usefulness. Granted, it is in the early stages yet, but by the end of the week we hope to have somthing useful (such as a raytracer) running using HPVM.
Timothy Baldridge
HPVM Project Manager
How are they useless?? I can find a lot of uses for an SGI Indy or Indigo that doesn’t require parallel computing. In fact, now that I know they’re that cheap on ebay, I’m going to pick one up! I’ve always wanted an SGI system.
So what’s the difference between this and PVM??
http://www.csm.ornl.gov/pvm/pvm_home.html
Parallel programming is usually harder… and if you have [a] “program that performs a repeated task on a large amount of data”, why not just submit the jobs to Gridengine (SGE) and let SGE do the scheduling.
SGE supports almost every platform, just download ths source and compile on your box:
http://gridengine.sunsource.net
Seems to me that grindengine is just distributed computing where i beleive the attmept here is cross platform beowulf that is easy to use.
I can tell you we consider old technology virtually worthless. We’re trying to pack as much computing power as we can into our server room with minimum consumption of floorspace, electricity, and cooling resources as possible. Also, since a cluster of N nodes have N points of failure (especially for non-fault tolerant MPI applications) increasing the complexity of the cluster increases its chances for failure.
Personally I’m looking forward to when/if Apple adopts the Cell processor for the Xserve and IBM releases a Fortran compiler designed to leverage its massive parallelism. To me that would be the ideal solution…
I can tell you we consider old technology virtually worthless. We’re trying to pack as much computing power as we can into our server room with minimum consumption of floorspace, electricity, and cooling resources as possible.
Sure but can you scale your existing HPC by simply adding a different Platform/OS? Besides i think one of the goals of this project is to make HPC available in a simple form to anyone that may have an interest.
It is pretty much the same as with that Xgl thingy presented in http://www.osnews.com/comment.php?news_id=9656. If a concept or technology is old enough, someone will resell it as an innovation.
pica
Hello,
do you know the term “embarrassingly parallel”. Beside Mandelbrot or Julia set computations, ray tracing ist one of the most famous embarrassingly parallel problem. In the late 80s and early 90s Parsytec presented Mandelbrot set computations to impress the visitors at fairs. After “Jurassic Park” ray tracing became quite popular.
So, what about FEM codes as an example of non embarrassingly parallel codes? What about latencies, due to network technologie used or conversions (e.g. little/big endianess)?
pica
You could probably get more bang for your buck with single processor or dual systems. The PIII systems won’t add up to too much power.
For $800 you could get 4 Cheap PC’s from Walmart with a combined 6GHZ+ of power.
“one single unified API is created to unite all these systems”
– it been here for years, it’s called NetBSD.
There are several clustering solutions available for it.
HPVM does have a way for converting between big/little endian machines. Sure you can make a cluster of NetBSD machines, but that’s not the point.
What cause me to want to try this idea in the first place, was the fact that at my work we use mostly Win32 machines, the video dept. uses Macs and I personaly perfer Linux. There is (as far as I have seen) no program out that will let me call up a deamon (or client) on each machine and use those machines when they are not otherwise in use. As one person above mentioned, it’s cross-platform Beowulf.
Sure it could be faster doing it with straight PVM, or NetBSD, but hey, if I can use Macs, PCs, SUNs and SGIs all together on the same network, I’ll take it.
Timothy Baldridge
HPVM Project Manager
wwwteo.informatik.uni-rostock.de/ ~hunger/wswos/DCW_papers/CarstenPitz.DCW98.ps
Helmar Gebert did a really fantastic job building realising my ideas and bringing in own ideas in his diploma thesis. The prototype he wrote was much more than just a prove of concept.
Carsten
While rereading my old work, I stumbled across the eMail addresses. Both are no longer valid.
Well, at Pallas I worked on MPI implementations for commercial MPP systems. With Pallas I left the scientific computing arena and jumped into the business computing arena.
Greetings,
Carsten