?

Log in

A sixty-second talk about my research - Lindsey Kuper [entries|archive|friends|userinfo]
Lindsey Kuper

[ website | composition.al ]
[ userinfo | livejournal userinfo ]
[ archive | journal archive ]

A sixty-second talk about my research [Nov. 19th, 2012|12:03 am]
Lindsey Kuper
[Tags|]

(Update: I got the PLMW travel award; hooray! I also changed my talk slightly from the version appearing below, because, as they say, certain things have come to light. Stay tuned.)

I just applied for a scholarship to attend PLMW 2013 in Rome next January, which, if I'm accepted, will cover airfare, hotel, and registration fees for both PLMW itself and POPL for the rest of that week. It's an amazing deal for those of us in the US -- round-trip airfare alone would cost over $1000 for me -- and I hope all the students I know in the PL community here will apply!

As part of the PLMW festivities this time, they're running a one-minute madness event:

One minute madness is a fun event where you get to present your research in 1 minute. You can have 1 slide to help you. After one minute you must leave the floor (escorted if necessary) and the next presenter begins.

The idea of presenting one's research in one minute is, well, madness. But I came up with the following talk, which clocks in at fifty-eight seconds:

Hi, my name is Lindsey Kuper, and I'm a Ph.D. student at Indiana University, where I'm working with Ryan Newton on new models for deterministic parallel programming. We're excited about deterministic parallel models because they offer programmers the promise of freedom from subtle, hard-to-reproduce nondeterministic bugs that are the scourge of traditional parallel programming. We've observed that in existing deterministic parallel systems, from venerable ones like Kahn process networks to modern ones like Intel's Concurrent Collections framework, the determinism of the system is based on some notion of monotonicity. By taking monotonicity as our starting point, then, we can generalize existing models. In particular, we generalize existing single-assignment models to allow multiple assignments that are monotonically increasing with respect to a user-specified partial order. We maintain determinism by allowing only monotonic writes and "threshold" reads to and from shared data. If you're interested in learning more, you can find our draft paper and tech report by doing a web search for "lattice-based deterministic parallelism". Thanks!

What do I mean by "traditional parallel programming"? That's a good question, but what I think I mean is that traditional languages only offered programmers one set of abstractions for doing either parallelism (that is, making sequential programs go faster by spreading some parts of the work across multiple available processors) or concurrency (that is, a program-structuring technique that allows multiple threads of control, for times when you have multiple distinct tasks to accomplish), and that single set of abstractions (threads and locks, or synchronized) was better suited for concurrency than it was for parallelism. If you wanted to do parallelism, you were forced to use a fairly unsuitable abstraction that allowed nondeterminism in places you didn't want it. So a deterministic parallel model, then, is really just one that recognizes that parallelism calls for a different set of abstractions than concurrency does.

I don't find there to be a shortage of deterministic parallel models in the research community -- quite the opposite. Our research isn't just about Yet Another deterministic parallel model; it's about sussing out the essence of deterministic parallelism. Having said that, our monotonic-writes-and-threshold-reads model is a twist that might be especially well suited for certain kinds of problems. We'll see!

LinkReply

Comments:
[User Picture]From: bubblingbeebles
2012-11-19 06:48 am (UTC)
this sounds like a lot of fun. i wonder how much variance there is among how long it takes to explain the groundwork/context for one's work. i expect probably the fun part of this is that you get to avoid talking about the tradeoffs your researches make and just mention the benefits.
(Reply) (Thread)
[User Picture]From: lindseykuper
2012-11-19 07:06 am (UTC)
Consider applying!

You also get to show a slide -- but only one. I suspect a lot of people's explanations would be helped by a 60-second video or animation of some kind. But that's in my fantasy world where we're all expert animators and illustrators.
(Reply) (Parent) (Thread)
[User Picture]From: mindstalk
2012-11-19 07:41 am (UTC)
Also depends on what the audience knows, and what you think they know, and how good you are at modeling that.

--guy who just re-read an old paper of his which veered from "let's explain groups to naive laymen" to "blah blah highly abstract technical stuff apparently aimed at his group theory classmates"
(Reply) (Parent) (Thread)
[User Picture]From: jcreed
2012-11-19 11:17 am (UTC)
Oh that sounds really cool! I didn't know you were working on that kind of thing. Monotonicity over user-defined orders sounds like a nice powerful thing to work with.
(Reply) (Thread)
[User Picture]From: jcreed
2012-11-19 11:18 am (UTC)
Also the "one-minute madness" link is giving me a 403, dunno why.
(Reply) (Parent) (Thread)
[User Picture]From: lindseykuper
2012-11-19 07:20 pm (UTC)
Dunno; works for me.
(Reply) (Parent) (Thread)
[User Picture]From: lindseykuper
2012-11-19 07:11 pm (UTC)
Yeah! I'd be curious about your thoughts, if you ever feel like reading about our stuff. We're working on getting a paper published now.
(Reply) (Parent) (Thread)
[User Picture]From: gwillen
2012-12-13 08:28 am (UTC)
I totally recommend lindseykuper's paper -- it is very accessible, by which I mean I was able to understand it without difficulty. ;-) I would be happy to discuss it with you after.
(Reply) (Parent) (Thread)
[User Picture]From: lindseykuper
2012-12-15 06:12 am (UTC)
Thanks, gwillen! I would, of course, be happy to join any discussions that you and jcreed have.
(Reply) (Parent) (Thread)
(Deleted comment)
[User Picture]From: lindseykuper
2012-11-20 07:00 am (UTC)
Hm -- actually, "We're excited about deterministic parallel models because" used to be "Deterministic parallel models are great because", but I changed it because it's up to other people to decide what they think is great! In any case, though, I think I intended the "we" less as "my advisor and me" than "a number of people in the research community".

Also, although I do think it's cool that deterministic parallel models let you have parallelism without throwing the determinism baby out with the bathwater, it's honestly not the most exciting part of the project for me. For me, the exciting part is stuff like how, when I did the determinism proof, a frame property just sprang up out of nowhere.

Edited at 2012-11-20 04:15 pm (UTC)
(Reply) (Parent) (Thread)
[User Picture]From: sajith
2012-11-21 08:22 pm (UTC)
Whoa, it's an excellent one-minute talk!
(Reply) (Thread)
[User Picture]From: lindseykuper
2012-11-21 08:29 pm (UTC)
Did you apply for the PLMW scholarship yet? *expectant look*
(Reply) (Parent) (Thread)
[User Picture]From: sajith
2012-11-21 11:58 pm (UTC)
It's open in another tab! I've been thinking about the longer sentences ("summarize your research experience", "what do you hope to gain from this workshop"), then I got distracted and stepped out into the nice weather. I'll get back to it shortly.
(Reply) (Parent) (Thread)