Lambda the Ultimate

I enjoyed the paper. The idea of intermediate keys on map output being efficiently partitioned into parallel reduce steps was nice, and the implementation (especially wrt fault tolerance and locality) was interesting

The paper mentioned application to machine learning. My question is: how does the framework handle multiple passes over the same data (say, for estimation-maximization)? Presumably, you could save yourself from reprocessing the input and simply reuse the same plan repeatedly (modulo failures) - the only thing that changes, and thus the only subsequent communication that the map nodes would need, is the previous iteration parameters (the reduce output).

Do you have the abstraction of a loop with reduce output becoming (part of) the next iteration's input, and an optimal dataflow for it? Of course, even being able to quickly perform one iteration would be nice.

Data-parallel programming is definitely back on the rise, at least on the (admittedly narrow) area of general computation using GPU's. If you enjoy seeing an architecture being turned inside out and being used for a completely different purpose, you should check some of these papers out. I suggest the SIGGRAPH 03 and Graphics Hardware 03 papers (with a nice index available online). Look for the Computation on GPU section on SIGGRAPH and the Simulation and Computation section on Graphics Hardware. With these consistent reports of significative speedups (usually you get 5-10x faster on a GPU, compared to similarly costed CPUs), I'd say that this is not something that is about to go away anytime soon. My Navier-Stokes solver can be 20x faster than a single-cpu solution </plug> :)

And to stick the obligatory language design link here, check out Brook. They call it stream programming, but it's just another name for data-parallel programming. In one of Pat Hanrahan's talks, he describes programming on GPUs as having to relearn most of the stuff the folks on the Connection Machines used. Pretty cool stuff if you ask me.

I'm pretty ignorant about graphics cards. Do you happen to know if any of these compilers work with an ATI Radeon Mobility M6 LY (as found in Thinkpads)?

Unfortunately, no. Pixel and vertex shader support only starts to show up in the Mobility Radeon line starting with the 9000 series. Time to get a PowerBook, all of which support pixel and vertex shaders, with nVidia's GeForce FX Go5200 on the 12" models, and ATI's Mobility Radeon 9700 on the others. :-)

...get some GPU with Shader Model 3 support (giving you dynamic branches and less restrictions on the shader code than SM1.1 and SM2), such as a NV4x ;)

Data parallel programming is also alive at Ab Initio Software Corp. with unfortunately little detail of the proprietary programming tools.

Manuel Chakravarty and Gabriele Keller are working on the the Nepal Project at the University of New South Wales.

Nepal takes elements from Nesl and Sisal and centers around array fusion.



Personally, I think that data parallelism will be much more important now that multi-core CPUs are in the pipeline.

Shae Erisson - www.ScannedInAvian.org

The Lispers seem to have more of a rough trott by and large, with a bit of concealing, apologising, and dressing-up-in-XML. Hopefully that's changed in the past year or two though.

(Nice to see you here by the way, Thomas! See you thursday?)

Oh, I've had to reboot (turn on & off) a cellphone now and again--plus hardware problems, of course.

It's not MS's fault of course

One should be very careful when attributing intention to MS. They only care about market share and profits, not about reliability, usability, security, or any technical issue. It is an amusing exercise to deconstruct all remarks by Bill Gates, Steve Ballmer, and so forth, regarding their effect on market share. It's rather obvious that no statement from MS should ever be taken at face value as meaning what it seems to mean. An encouraging corollary is that MS is actually a company that is very consistent in its behavior and easy to understand and predict, much more so than companies that are more technically oriented. Almost nobody predicted Apple's coup in the legal Internet music business. Almost everybody predicts correctly that Microsoft instantly created a long-term goal to monopolize that particular area.

Ummm. OK. Mod up (Insightful).

<wink>!

Plenty of people are touched every day by AI programs
- Search engines (via Naive Bayes Classifiers)
- OCR (Neural nets)

That's two. Then there are constraint processing engines, like those that deal with autoreserving airline tickets. Also, those new speech-recognition automated telephone menus. Oh how I hate those things, but that's AI: Language Models.

It's not AI, it's that no-one wants to claim AI. It reeks of "big brother" and the End Of The World, because of all the bad press from the buzzword loving mainstream media. Things like Terminator 1-2-3, Colossus, 2001, and The Matrix all claim that AI will destroy the world.

As for more people having been touched by AI than telephones -- probably true, but not as drastically as such a broad statement would lead one to believe.

We are talking about customer perception here. On an everyday level, people tend think in terms of objects that they can touch, not a process, which is more abstract.

For instance, people say "best thing since sliced bread", and not "best thing since bread slicers", or "best thing since bread slicing machines", or "best thing since bread slicing".

AI is the slicing, while telephone is the sliced bread.