Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

research.sun.com

Post reply on HN
Print view View original

11–20 of 31 posts

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#11

Post by gruseom »

tsally wrote: post #7

Clojure tries to address this problem. First and rest are still fundamental concepts in its data structures, but underneath the hood is properly written concurrent Java code. Hickey wrote it once, and now you don't have to. I highly recommend the Clojure videos, available at: http://clojure.blip.tv/ . Watch them online or download them via iTunes. There are a two videos on Clojure data structures, and a set of introd…

Hickey wrote it once, and now you don't have to.

But Steele is talking about introducing fundamental new constructs all the way up to the application level. This is not a problem that can be abstracted away by better language implementations, and certainly not by better implementations of familiar constructs.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#12

Post by silverlake »

Bring back SETL. Higher order operations on abstract sets can hide the complexity of parallelism. Under the hood you can implement sets using whatever data structure fits your hardware (single core, multiple cores, GPUs, weird memory hierarchies).

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#13

Post by apgwoz »

gruseom wrote: post #11
tsally wrote: post #7

Clojure tries to address this problem. First and rest are still fundamental concepts in its data structures, but underneath the hood is properly written concurrent Java code. Hickey wrote it once, and now you don't have to. I highly recommend the Clojure videos, available at: http://clojure.blip.tv/ . Watch them online or download them via iTunes. There are a two videos on Clojure data structures, and a set of introd…

Hickey wrote it once, and now you don't have to. But Steele is talking about introducing fundamental new constructs all the way up to the application level. This is not a problem that can be abstracted away by better language implementations, and certainly not by better implementations of familiar constructs.

> and certainly not by better implementations of familiar constructs.

Why not? I sort of took the slides as just that. For a good chunk of the slides he talks about the simple singly-linked list found in LISP. Then shows that implemented as a Tree (conc list), familiar operations still exist, but we get improvements by making operations more parallelizable.

Slide 24 starts showing how `first`, `rest`, `append` can be implemented with conc lists, and 31 shows `map`, `reduce`, `mapreduce` as well. Then, he goes on to show `length`, `filter` and `reverse` in terms of `mapreduce`.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#14

Post by gruseom »

apgwoz wrote: post #13
gruseom wrote: post #11

Earlier quoted context omitted.

Hickey wrote it once, and now you don't have to. But Steele is talking about introducing fundamental new constructs all the way up to the application level. This is not a problem that can be abstracted away by better language implementations, and certainly not by better implementations of familiar constructs.

> and certainly not by better implementations of familiar constructs. Why not? I sort of took the slides as just that. For a good chunk of the slides he talks about the simple singly-linked list found in LISP. Then shows that implemented as a Tree (conc list), familiar operations still exist, but we get improvements by making operations more parallelizable. Slide 24 starts showing how `first`, `rest`, `append` can be…

Why not?

Well, if first and rest are unsuitable for parallelism as Steele says, a programming edifice built on top of better implementations of those might be beside the point.

My understanding of the slides is that we don't need better language runtimes to exploit parallelism for us while we continue to write the same kind of code as before; what we need is a new set of primitives (modeled by conc lists in this talk) that would allow us to compose programs differently, with more parallelism. And that far from being insulated from this, programmers will have to be thinking about it throughout.

It's possible that Clojure's implementation is already such that these primitives, and a programming style built on them, are feasible today. If so, that's awesome. But this is a different thing than providing better implementations of cons-based abstractions.

Edit: another thing - the transformation from cons to conc is nontrivial in the Lisp context because Lisp's code=data is so tied to cons. It might be an interesting to think about what programs would look like if you applied code=data rigorously in a conc model. Would a language whose expressions are represented as concs look or feel significantly different than a cons-based one?

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#15

Post by dreish »

tsally wrote: post #7

Clojure tries to address this problem. First and rest are still fundamental concepts in its data structures, but underneath the hood is properly written concurrent Java code. Hickey wrote it once, and now you don't have to. I highly recommend the Clojure videos, available at: http://clojure.blip.tv/ . Watch them online or download them via iTunes. There are a two videos on Clojure data structures, and a set of introd…

This is confusing the difference between concurrency and parallelism, the latter of which Clojure doesn't yet address.

If your program is divided into threads, having a language like Clojure that makes concurrency easy and lock-free is a wonderful thing, and of course you get some benefit from running on multiple cores, though I don't think that's entirely the point.

If, on the other hand, you're looking for a faster way to map a function across eight million lines in a flat file, using Clojure's map function won't magically make your program run faster (and pmap might not run at all -- you may run out of memory, since pmap isn't truly lazy).

The fact that a Clojure PersistentVector (not a PersistentList) is implemented with a 32-way tree (not 64-way) does not imply that Clojure has abandoned the car/cdr paradigm in favor of conc lists -- it still uses cons, in a more generic form, but still with the same first/rest concept. The trees allow vectors to be inexpensively persistent, so you can have actual practical immutable vectors.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#16

Post by drobilla »

mgreenbe wrote: post #9
giardini wrote: post #8

Earlier quoted context omitted.

That would be OK if Lisp wasn't list-based. In Lisp you always start at the beginning of the list: any processing requires going through that list somehow. Prolog allows some workaround to this problem by using difference lists.

Did you not read the talk? You can provide the car/cdr interface on top of tree-based implementations. Yes, you pay a cost to use that interface...but you pay a different (Guy Steele would argue more expensive) cost if you write your algorithms using car/cdr instead of, say, empty/singleton/conc . Moreover, as is pointed out on the slides , a lot of processing can happen without ever having to talk about car/cdr , e.…

This aspect of the talk is a big justification for more abstraction in a lisp (namely interfaces / type classes / similar).

This decision should not actually affect written code, other than maybe a type declaration or similar to say "this is a singly linked list" or "this is a tree list". Even if you ignore performance, that car really means "get the first element of this two element cons cell" rather than "get the first element of this sequence" is one of the most obvious limitations / warts of (classical) lisp (and scheme) as compared to more modern languages designed with data abstraction in mind.

Of course, this is "modern" as in OO decades ago (Java, C++, etc). Taking it a step further you could omit the type declaration entirely and infer at compile or runtime (via profiling + JIT) which list implementation is best.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#17

Post by tsally »

dreish wrote: post #15
tsally wrote: post #7

Clojure tries to address this problem. First and rest are still fundamental concepts in its data structures, but underneath the hood is properly written concurrent Java code. Hickey wrote it once, and now you don't have to. I highly recommend the Clojure videos, available at: http://clojure.blip.tv/ . Watch them online or download them via iTunes. There are a two videos on Clojure data structures, and a set of introd…

This is confusing the difference between concurrency and parallelism, the latter of which Clojure doesn't yet address. If your program is divided into threads, having a language like Clojure that makes concurrency easy and lock-free is a wonderful thing, and of course you get some benefit from running on multiple cores, though I don't think that's entirely the point. If, on the other hand, you're looking for a faster…

Ah yes, a silly mistake on my part. I am hoping that eventually Clojure will help the programmer write parallel algorithms, but you are correct in that such constructs don't exist in the language at this moment.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#18

Post by mquander »

gruseom wrote: post #14
apgwoz wrote: post #13

Earlier quoted context omitted.

> and certainly not by better implementations of familiar constructs. Why not? I sort of took the slides as just that. For a good chunk of the slides he talks about the simple singly-linked list found in LISP. Then shows that implemented as a Tree (conc list), familiar operations still exist, but we get improvements by making operations more parallelizable. Slide 24 starts showing how `first`, `rest`, `append` can be…

Why not? Well, if first and rest are unsuitable for parallelism as Steele says, a programming edifice built on top of better implementations of those might be beside the point. My understanding of the slides is that we don't need better language runtimes to exploit parallelism for us while we continue to write the same kind of code as before; what we need is a new set of primitives (modeled by conc lists in this talk…

> It might be an interesting to think about what programs would look like if you applied code=data rigorously in a conc model. Would a language whose expressions are represented as concs look or feel significantly different than a cons-based one?

Now that is food for thought.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#19

Post by herdrick »

I'm glad someone is working on this but I think parallel 'map' and 'filter' (which is 'easy' for compiler writers) will be plenty for the next five years. And how sure is anyone about the future of hardware beyond that?

Still this is really interesting. Can't wait to see the video.

Re: ICFP '09: "Get rid of cons!" Guy Steele on parallel algorithms & data structures

#20

Post by herdrick »

gruseom wrote: post #14
apgwoz wrote: post #13

Earlier quoted context omitted.

> and certainly not by better implementations of familiar constructs. Why not? I sort of took the slides as just that. For a good chunk of the slides he talks about the simple singly-linked list found in LISP. Then shows that implemented as a Tree (conc list), familiar operations still exist, but we get improvements by making operations more parallelizable. Slide 24 starts showing how `first`, `rest`, `append` can be…

Why not? Well, if first and rest are unsuitable for parallelism as Steele says, a programming edifice built on top of better implementations of those might be beside the point. My understanding of the slides is that we don't need better language runtimes to exploit parallelism for us while we continue to write the same kind of code as before; what we need is a new set of primitives (modeled by conc lists in this talk…

I don't think so. Lisp code is already a tree.
Post reply on HN