David Patterson Says It’s Time for New Computer Architectures and Languages • Hacker News Board

Re: David Patterson Says It’s Time for New Computer Architectures and Languages

#201

Post by sitkack » Wed, Sep 19, 2018, 6:04 PM UTC

eslaught wrote: post #198

sitkack wrote: post #155

Earlier quoted context omitted.

How does Legion differ from languages like Halide which separate scheduling from the algorithm? How much of this could get handled by PGO (profile guided optimization) to observe locality and access patterns? Do you think AI has a place in directing scheduling given enough semantic information so that "bad hypothesis" can get pruned? Does it rely on a static description of machine (latencies and bandwidth) or does it…

Wow, this is a tough set of questions! Some initial thoughts: > How does Legion differ from languages like Halide which separate scheduling from the algorithm? Yes, that's one of the similarities. Legion, Sequoia, and Halide all share a separation between the specification of the algorithm from the mapping to any particular architecture. The biggest difference between Legion and Halide, aside from the fact that Halid…

Thank you for such a detailed response, I am grateful.

You mentioned Terra which is an inspiration of mine, I really enjoy mixing high level and low level code and being able to do so in the same program which is one of the major affordances of LuaJIT.

Hugo M. Gualandi and Roberto Ierusalimschy have released the core Lua teams' version of Terra, in Pallene [1] [2] which has the austere aesthetic of Lua applied to ideas of Terra. Pallene looks to be the next incarnation of Titan [3].

[1] http://www.inf.puc-rio.br/~roberto/docs/pallene-sblp.pdf

[2] https://github.com/pallene-lang/pallene

[3] https://github.com/titan-lang/titan

Re: David Patterson Says It’s Time for New Computer Architectures and Languages

#202

Post by Liskni_si » Wed, Sep 19, 2018, 10:25 PM UTC

jayd16 wrote: post #149

tabtab wrote: post #28

I have to say that most bottlenecks are from lazy designs, not wimpy hardware. As a thought experiment, suppose parallel processing scaled easily without worrying about Amdahl's Law and similar bottlenecks. So then we put 128 cores into our computers. Software companies would eat up every single core eventually if there's no penalty for inefficient coding. They don't pay our electric bill, we do. It's kind of like fr…

>It's kind of like freeways: the more you build, the more the traffic increases to fill them up again. I see this a lot. Is this a real phenomenon? Seems like an inevitability to me. Who would build a freeway they didn't expect people to use?

Yes, see https://en.wikipedia.org/wiki/Induced_demand

Re: David Patterson Says It’s Time for New Computer Architectures and Languages

#203

Post by uryga » Fri, Sep 21, 2018, 12:45 AM UTC

alexisread wrote: post #185

uryga wrote: post #178

Earlier quoted context omitted.

I'm very curious about your solution! If there's a place where I could follow the development, I'd be grateful for a link or, if it's not too much of a hassle, a ping to the email in my profile when you have something you want to show off :)

I'm grateful for the interest! Not really thought about a location - I'll probably use my github account when I've got some material, and ping you when it's there. As an aside, most of the stuff I've on my account is higher level eg. https://github.com/alexisread/noreml deals more with a new method of visually programming cross-plaform UIs ie. using nodered dashboards and flow-based programming in a drag and drop fas…

Thanks, and good luck! Followed you on GH. noreml sounds interesting too, I'll check it out when I have better internet access!

Re: David Patterson Says It’s Time for New Computer Architectures and Languages

#204

Post by noahdesu » Fri, Sep 21, 2018, 3:44 PM UTC

eslaught wrote: post #189

eslaught wrote: post #117

Earlier quoted context omitted.

(Disclaimer: I work on Legion) Sequoia has been largely superseded by its spiritual successor Legion [1], another programming system by Alex Aiken. Legion doesn't focus so much on low-level memory hierarchies, but it does a very, very good job of scaling to very large machines, and taking advantage of heterogeneous processors such as GPUs. It is also incomparably better at working with dynamic programs, whereas Sequo…

I'm not sure if anyone is still reading this thread, but I wrote up some thoughts on the differences between Sequoia and Legion and some of the lessons we learned. Maybe this is a little too much "inside baseball", but hopefully some will enjoy it. Sequoia was basically an answer to the question, "what could a compiler do for you if it knew everything about your program?" And I do mean everything : in addition to you…

Thanks for this write-up. I can't wait to see what comes out of the Legion projects next. One of the most exciting things I think that is going on.