Inside Intel's $20M Multicore Research Program

An anonymous reader writes "You may have heard about Intel's and Microsoft's efforts to finally get multi-core programming into gear so that there actually will be a developer who can program all those fancy new multicore processors, which may have dozens of core on one chip within a few years. TG Daily has an interesting article about the project, written by one of the researchers. It looks like there is a lot of excitement around the opportunity to create a new generation of development tools. Let's hope that we will soon see software that can exploit those 16+core babies. 'The problem of multi-core programming is staring at us right now. I am not sure what Intel's and Microsoft's expectations are, but it is quite possible that they are in fact looking at fundamental results from the academic centers to leverage their large work force to polish and realize the ideas that come forth. It calls for a much closer collaboration between the centers and the companies than it appears at first sight.'"

Comments

It's easy

[Anonymous Coward]

./configure --num-cores=16

Will more cores help me decipher this run-on?

[alta]

I am not sure what Intel's and Microsoft's expectations are, but it is quite possible that they are in fact looking at fundamental results from the academic centers to leverage their large work force to polish and realize the ideas that come forth. Maybe my brain needs a new compiler. This must be a multi-core sentence.

Show me the money Intel.

[stratjakt]

SMT processors of this type are only useful for accelerating a certain type of problem set, and useless for most general computing.

We've had SIMD multicore PC's forever, and they're useless as desktops. I write this from a quad xeon machine, repurposed as my dev box, as CPU1 grinds away at about 75% all day long, the rest idle. It's been like that for more than a decade, it'll be like that until MIMD hits the street with a whole new paradigm of programming languages behind it - a handful of C compiler #pragma directives from intel isn't going to make this work.

It's not simply a matter of "coders don't know how to do it." It's a matter of these multi-core "general purpose" CPUs are only really useful for a fairly limited set of specific problems.

Eg; writing a game engine with a video thread, audio thread and an input thread still leaves 13 cores idle. You really cant thread those much farther (the ridiculously parallel problem of rendering is handled by the GPU).

Simply starting processes on different procs doesn't help all that much, since they all fight over memory and I/O time. The point of diminishing returns is reached fairly quickly.

But hey, if all you do is run Folding@home so you can compare your e-cock with the other kids on hardextremeoverclockermegahackers.com, well I have some good news!

As for me, I'm seeing AMD's multiple specific purpose core approach as being more viable, as far as actually making my next desktop computer perform faster.

Savain says it best at rebelscience.org: "Even after decades of research and hundreds of millions of dollars spent on making multithreaded programming easier, threaded applications are still a pain in the ass to write."

Re:Show me the money Intel.

[everphilski]

a handful of C compiler #pragma directives from intel isn't going to make this work.

That's OpenMP, and depending on the program, it can work wonders. In an hour I parallelized 90% of a finite element CFD code with it. Yes, it sucks for fine-grained parallelization.

Intel's product is Threaded Building Blocks, and is not built around pragmas, and is both commercial and OSS. It's pretty slick and will let you do the more fine-grained optimizations.

It's a matter of these multi-core "general purpose" CPUs are only really useful for a fairly limited set of specific problems.

Not entirely true, it's just useful for problems that need a processor.

I write this from a quad xeon machine, repurposed as my dev box, as CPU1 grinds away at about 75% all day long, the rest idle.

... obviously, you have more processor than you need. I, on the other hand, have a quad core Opteron that is currently over 350% utilization. I tank it almost 24/7.

the ridiculously parallel problem of rendering is handled by the GPU

Not for long. Raytracing is making a comeback.

As for me, I'm seeing AMD's multiple specific purpose core approach as being more viable, as far as actually making my next desktop computer perform faster.

If you can't even tank one core of your Xenon, it's doubtful.

"Even after decades of research and hundreds of millions of dollars spent on making multithreaded programming easier, threaded applications are still a pain in the ass to write."

I'd caveat that by saying "threading arbitrary program X is a pain in the ass." There are plenty of useful programs that are easily parallelized.

Re:Most PCs are fast enough

[pla]

Unless the next version of Windows requires a core dedicated to the OS or something in the future.

So, uh, you haven't Vista yet, I see...

Re:Most PCs are fast enough

[KillerCow]

The thing is, most PCs have plenty of computing power as a single core system

And 640k ought to be enough for anyone.

I think as time moves on and quad core becomes the "low-end" you will see less demand for higher end hardware.

My last purchase (6 to 8 months ago) was a "low-end" machine. I chose carefully to make sure that it was low-end and not bargain-basement. It has two cores. I don't think it's even possible to buy a single core machine through mainstream channels anymore. Today's low-end (multi-core) is more than adequate for most users to use over the next few (read: four) years.

Unless the next version of Windows requires a core dedicated to the OS or something in the future.

You do not understand how the scheduler works.

Moving the bottleneck...

[MarkEst1973]

Forget software not being written for multi-cores, the entire infrastructure around the computer needs to "go wide" for massive parallelism, not just the software. This includes disk, memory, front-side bus, etc./p>

I'm doing highly concurrent projects (grid computing) for my company and we're finding that some things parallelize just fine, but others simply move the pain and bottleneck to a piece of infrastructure that hasn't quite caught up yet.

For example, my laptop has a dual-core 2.2Ghz processor, which you'd think is great for development. It's no better than a single CPU machine because my disk IO light is on all the time. IntelliJ pounds the disk. Maven and Ant pound the disk. Outlook pounds the disk. Even surfing the web puts pages into disk cache, so browsing while building a project is slow. Until I get a SCSI drive, you're still limited on disk IO, so those extra cores don't help that much.

All the cores are great on the server, though. I've recently completed a massive integration project where I grid-enabled my company's enterprise apps. All those cores running grid nodes is giving us very high throughput. Our next bottleneck is the database (all those extra grid nodes pounding away at another bottleneck resource...)

Terracotta Server as a Message Bus. It's been a very interesting project.

Re:stupid much?

[Jerry Coffin]

Instead of trying to convince everyone on Earth to change all existing software, why doesn't Microsoft just make the next version of Windows have a process handler that can process single threads on multiple cores at once? Actually technically I think Intel could do that internally on their processors too sort of like RAID for cores.

Intel's been doing that (to some degree) since the Pentium, and they increased it a lot in the Pentium Pro/Pentium II. It works reasonably well up to a point (modern chips typically execute an average of two instructions per clock cycle) but definitely has limits.

Compilers to automatically detect when instructions can be executed in parallel have been around for years. Cray had vectorizing compilers by the late 1970's, and within rather specific limits, they worked perfectly well. Just for example, if you wrote a loop like:

for (int i=0; i<256; i++)
a[i] = b[i] * c[i];

they'd break the loop down into four actual executions of a loop, each of which worked on 64 items in parallel. It had independent execution units, so at a given time it'd normally be loading one set of 64 items into one set of registers, executing multiplications on a second set of 64 items, and storing results from a third set of 64 registers.

That has a couple of problems though. First of all, if you're not careful, it's pretty easy to create loops with (apparent) dependencies from one iteration to the next, so the compiler can't parallelize the code. Second, this works well for vector processors, but probably not nearly so well for a large number of completely independent processors (which have higher communication overhead, meaning that starting up things to happen in parallel is more expensive).

If you're willing to provide the compiler with a little help, it can do quite a bit more, such as with MPI. The standard MPI interface is pretty low-level, but if you want to do the job in C++, Boost.MPI helps out quite a bit (cheap plug: if you want to know more, consider attending Boostcon '08).