Computer Architecture is Back: Parallel Computing Landscape

[joeybonesgnarly]

this dude seems nervous, pulled through and made a great presentation though so props

[iampeeay01]

Thanks so much for sharing this video. learned a lot from it.

[thegreeensky]

nice video.. nice discussion...

[grunder20]

Merry Christmas everyone! Good work on this video!!

[insomniacgrace]

Thanks for having this video posted. Nice video clip. A great help for everyone whose into this field.

[agapitoflores001]

This video is excellent. I'll be waiting for more videos like this.

[Kamikaze576]

Rate up if you are being forced to watch this for ECE 222 at the University of Waterloo!

The University of Waterloo likes to steal lectures from other universities to make their courses less crappy.

[HostileHST]

@Kamikaze576 It's a business, what did you expect, they do whatever they can to be cheap, seems like a lot of work to me, lol

[flowewritharoma]

Parallel Computing Landscape is intaractive!!! old-type straight computer has limited by theory of tempurature and size. So we must think a new-type computer with cloud/quntamu.

[dagod79]

"I don't care about Word............I care about the future!".

See, the thing is, that your Future Word will be controlled via your voice, will do multi-language auto-correction and translation and "predictive auto-completion" by comparing your text/sentence to a billion of stored "preset" phrases/sentences/passages/tex­ts and automatically offer you endings for your phrase/sentence/or even complete paragraph/text. And this will benefit from parallelism just as much as any other program.

[evanplaice]

As far as the many-core vs multi-core argument goes. It depends. For system architectures that accomplish a simple canned set of tasks for a large number of requests and contain dispatchers capable of managing that system effectively, many-core is a clear winner. Examples include, servers and/or databases where the tasks are simple but the scalability isn't. For general purpose operating systems, unless they evolve to run hundreds of processes, multi-core is still the better choice.

[evanplaice]

I also wanted to point out that the heterogeneous model of computing already exists. The slow cores are represented by the CPU (and it's cores/processors) and the fast core used for complex number crunching is represented by the GPU. If you look at the structure of a modern GPU it looks similar to a miniture of the system it's attached to with its own memory, cache, data bus, etc, just a lot more specialized.

[evanplaice]

If you look at the software architecture of Google Chrome, there is an executive process that manages the application-wide state information and subsequent processes for each component in the application. That way, if one tab (or component of the application) crashes Chrome won't also be taken down with it. If you look at the organization it isn't a whole lot different than running a server farm. Except, the communication barrier becomes cross-application instead of cross-host.

[evanplaice]

I'd like to point out a few things. First, I'm not sure why the professor mentions using threading over multi-process programming architecture. Threading, while being easier to implement, is inherently more unstable than using multiple processes. First, the program and it's threads must share the same memory so, even if they're executing on different cores the cache still needs to be copied across (for a performance hit). Second, if a thread crashes it takes down the application.

[eternal000sunshine]

Geenius!!!

[JesusWASdarkSkin]

Is Dave Patterson is YVR authority now

[sean0101a]

I think using parallel computing is more about choosing the right algorithms and datastructures, we can do that with many existing languages.

[dagod79]

@sean0101a We will have to give away to the machine. Let it waste cycles to come up with new concepts by making MISTAKES. Let it evolve by experience. Let us call it "Computational Genetics" and let the machine reprogram its own genes. Actually, there's no other way for it to work in the long run. Yes, Star Trek (The Next Generation) and Data are our "blue prints". We need new languages and completely revolutionary compiler concepts in order to profit from the Core-Revolution.

[randomlaughingman]

My comment keeps being censored off here. The gist of it is that, it'd work to use thread-scheduling types of algorithms to split up non-parallel programs and run them in parallel environments (whether that's clusters, multicores, multiple processers).

Come up with an abstract works-for-anything outline. Kind of a mix of pipelining and job scheduling. Yes? No? Why no if so?

[sonnyrao]

Yeah, good video, the things that bother me are the hand waving at the end about how to bridge traditional clunky apps to this new model, i'm not so sure the traditional poorly written software is going to simply disappear anytime in the next 15 years

So I think the work is good, ideas are good for the most part, but its going to be up to "industry" to figure out how to actually bridge between old and new if research don't want to tackle that problem because its too hard.

[gaochao121212]

so interesting lesson

[wazborn123]

Excellent lecture

[colinmaharaj]

Excellent view of the past and how and why the present and the future exists the way it does (and will).

It is almost like the hardware gals/guys pushes the promise (problem) of Moores law to the software gals/guys.