Technology Brief for Apple's Grand Central Multicore Paradigm [pdf] • Hacker News Board

Re: Technology Brief for Apple's Grand Central Multicore Paradigm [pdf]

#11

Post by Gary_W_Longsine » Tue, Jun 09, 2009, 5:16 AM UTC

gchpaco wrote: post #4

This encourages a very different style of multicore programming than is common on desktops today, closer to that of grid systems. Very interesting to see how this develops.

Apple may have been inspired in part by their experience with XGrid [http://www.apple.com/server/macosx/technology/xgrid.html], based on the NeXTSTEP application Zilla [http://www.omnigroup.com/mailman/archive/macosx-talk/2000-Se...], and perhaps by PDO (Portable Distributed Objects) [http://en.wikipedia.org/wiki/NeXTSTEP]. I mention these somewhat obscure influences in response to your grid observation. Obviously there are several other (probably more immediate) influences on GCD and OpenCL as well (LLVM, OpenGL, various academic research on parallelized programming, etc.)

Re: Technology Brief for Apple's Grand Central Multicore Paradigm [pdf]

#12

Post by Gary_W_Longsine » Tue, Jun 09, 2009, 5:38 AM UTC

swannodette wrote: post #2

While all the Apple hardware announcements look great, I think this is possibly an even more significant announcement. C/Objective-C/C++ gets lambdas and closures! It looks like Apple has a very interesting single machine concurrency story here. Right now I'm really enjoying Clojure- it's lockless approach to single machine concurrency is plenty, plenty fun. Unfortunately the downside of learning Clojure is that many…

Related to this is another key performance technology in Snow Leopard, OpenCL: http://images.apple.com/macosx/technology/docs/OpenCL_TB_bri... .

Re: Technology Brief for Apple's Grand Central Multicore Paradigm [pdf]

#13

Post by Gary_W_Longsine » Tue, Jun 09, 2009, 5:50 AM UTC

agocke wrote: post #10

wmf wrote: post #6

Earlier quoted context omitted.

Obviously pthread support won't be removed; pthread programs just won't play as nice with each other as Grand Central programs.

Well I didn't really mean to suggest that it would be removed, mainly about what kind of performance hit I would take by using pthreads instead of Grand Central. In addition, how well will pthread programs and GC programs work together. Normally, the threading is handled by the scheduler in a straightforward model, but if the GC model modifies the scheduler model, I'm not really sure what the results will be.

You won't "take a hit", but you will "fail to utilize the hardware as efficiently as you could". Both GCD and OpenCL have their roots in open extensions to C. Both offer potentially significant performance gains for certain types of tasks. If enough developers find it advantageous to use either blocks or OpenCL, these technologies will likely show up on other platforms.

Regarding the scheduler, I'm not sure enough details are known outside of Apple at this point to know what the results would be on Snow Leopard. Your milage would obviously vary when porting to some other OS. It's easy to imagine that running a PThreads app might hog some cores, or be punished inappropriately by the scheduler. These things can happen when running more than one threaded app on a system today, without GCD. This behavior, even if not documented, should be relatively easy to explore with some simple demo apps, though, without porting your entire app.

And, uhm... you will need to refactor your application to take advantage of the new API. (Was that a trick question?) Apple's guidance indicates that you can make modest changes for big gains by focusing the refactoring effort on the parts of the app that do the most computationally. For some apps this can be a small amount of work for a large gain, particularly if you can take advantage of OpenCL as well as GCD.