Breaking News
Blog

Multicore Expo: Panel to challenge the state-of-the-art in parallelization

Markus Levy
4/15/2010 03:00 PM EDT

 2 comments   post a comment
NO RATINGS
View Comments: Newest First | Oldest First | Threaded View
Cozens
User Rank
Rookie
re: Multicore Expo: Panel to challenge the state-of-the-art in parallelization
Cozens   4/22/2010 12:06:19 PM
NO RATINGS
I don't know if I'm a maverick or even if what I'm proposing is any sort of silver bullet, but I have been working in a design for a "General Purpose Multiprocessor Operating System" that I call "Cooperating Operating Systems". I am documenting my ideas at http://greatc.wikia.com and I am making notes for a presentation at http://greatc.wikia.com/wiki/Presentation001 These notes aren't as complete as I would like because I've been doing this work in my spare time. I outline some of the ideas here. The design started by asking the question if there was an unlimited number of processors what would an operating system need to do? If we assume that a process runs on its own processor, then when a process wants to fork another process it needs to locate a free processor. It is the operating system's function to find an available processor. What is required is something like an employment agency, that for each free processor holds the processor's contact details (Processor Id). When a process forks it requests a processor Id from the OS and then communicates with this processor and launches the forked process. We want forking to be lightweight and fast, so for every application processor we attach a very small OS processor called a Q-Cell. Every Q-Cell holds a copy of the free processors. When a process forks, the Q-Cell associated with the process broadcasts a REQUEST message identifying the processor Id that will be used. If the broadcast is successful all Q-Cells update their processor lists and the processor Id is passed to the forking process. The Q-Cells communicate using a separate network to the inter process communication network. In a stable environment, that does not require acknowledgements, the time to fork a process is approximately the time to broadcast a message. If the rate of process forking is low compared to the time for a Q-Cell to broadcast, this scales well to 1,000+ processors. With a modification to the broadcast network structure it should be scalable to 1,000,000+ processors. This is only a very short summary of what I have been working on but I'd be interested to know if there is any interest in these ideas.

Mapou
User Rank
Rookie
re: Multicore Expo: Panel to challenge the state-of-the-art in parallelization
Mapou   4/15/2010 9:47:28 PM
NO RATINGS
"There will NEVER be a "silver bullet" to save programmers from the complexities of multicore‚??at least that's what industry experts believe." * Nah. You should have said, "that's what industry experts would like the rest of us to believe." Why? Simply because the big players in the multicore industry (whom you represent as president of the Multicore Association, Mr. Levy) have got way too much invested in last century's flawed paradigms and technologies to turn the boat around. However, this strategy of self-delusion can only work for a little while. Sooner or later, some maverick startup will pop up out of nowhere and show the industry the right way to do things. At that time, even the big dogs in the business will find themselves in a world of hurt. * What is really amazing is that the solution to the parallel programming crisis has been staring the industry in the face for decades. However, the members of the Multicore Association are too married to the past to see the writing on the wall: The Turing Computing model is not the solution but the problem. Of course, since the industry and the academic community are run by Turing Machine worshipers, they are stuck in a rut of their own making. That's too bad because the market does not care. The market wants a solution yesterday and will embrace anybody who comes out with it. * Google "How to Solve the Parallel Programming Crisis" if you are truly interested in finding a silver bullet. And not just for parallel programming but for software reliability as well.

Flash Poll
Like Us on Facebook

Datasheets.com Parts Search

185 million searchable parts
(please enter a part number or hit search to begin)
EE Times on Twitter
EE Times Twitter Feed
Top Comments of the Week