Log in

No account? Create an account
2012 in review (part 2 of 2) - Lindsey Kuper [entries|archive|friends|userinfo]
Lindsey Kuper

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

2012 in review (part 2 of 2) [Jan. 6th, 2013|09:03 pm]
Lindsey Kuper
[Tags|, , , ]

Continued from part 1!

As June came to an end, I went to back-to-back Mountain Goats concerts -- John Darnielle solo shows, really -- on the 27th and 28th, bringing Alex with me for the first (it was his first Mountain Goats show) and my friend Emily for the second (her first, as well). Because I lived in Mountain View, I spent twice as much time getting to and from these shows on public transit than I spent actually at the shows, but it was worth it, and I was really happy to get to give Alex and Emily a proper introduction to a band I've loved for a long time. Neither show quite left my head for months. (LMA has the first night's recording, which happened to be the night that JD played not one but two songs about professional wrestling.)

July arrived, and Ryan and I made a final push to finish up our POPL submission on time, with the help of many friends and colleagues who read drafts of it for us. Ken Shan spotted a couple of particularly subtle bugs.

Just before the deadline, a friend of Ryan's pointed us to some related work we hadn't seen yet: a very recent tech report from the group working on the Bloom language at Berkeley. Then, it turned out that my Mozilla colleague Niko had a friend who had worked on Bloom. Niko put me in touch with his friend, I started talking to the Bloom folks, various enthusiastic emails were exchanged, and I ended up being invited to go visit Berkeley and give a talk in August.

Things that Jim can help you with at OPLSS.

With the POPL deadline behind me, I took two weeks off from Mozilla to attend OPLSS in Eugene, Oregon, an annual two-week-long summer school that draws PL folks from around the world. The lectures were intense, mind-bending, and entertaining, and I loved hanging out with the people at OPLSS.

But it was at around this point in the summer that I began to feel pulled apart at the seams, because there were so many things I had to do and they all seemed critically urgent. Although our paper was now done, I had returned to trying to fix the remaining bug in our determinism proof, so that Ryan and I could put up the tech report that was supposed to accompany the paper. I felt guilty for not doing the OPLSS homework -- after attending the lectures all day, whatever energy I had left was going toward working on the proof instead. On the other hand, I felt equally guilty for being at OPLSS at all instead of back at Mozilla hacking on traits, which had gone rather quickly in my mind from being a somewhat exotic, nice-to-have feature to an essential feature that the language sorely needed. (Indeed, Niko's type inference code was filled with plaintive comments about how the implementation could have been so much cleaner, if only we had traits.) And on top of everything else, I had to prepare for my upcoming talk at Berkeley.

After a few days at OPLSS, with all my various research and work obligations at war for my attention, I was a stressed-out wreck. Alex listened to my woes on the phone with characteristic patience and understanding. Another thing that helped was getting out of my dreary, isolating OPLSS dorm room (my roommate had been a no-show, so it was just me) and running on the beautiful network of trails around Eugene. Nevertheless, with all the other stuff I had to do, I began to fall behind on the 20-mile-a-week habit I would have had to keep up in order to make my 1000-mile goal for the year.

I began to make progress on the proof again; then, at the start of the second week of OPLSS, I got a big emotional boost: Ryan and I got official notification that our grant had been funded! This was pretty huge news for me: it meant that my plans for doing a Ph.D. with Ryan were starting to shape up in the way that I had hoped.

By the end of OPLSS, I had managed to isolate and finally really understand the renaming bug. I still hadn't managed to fix it properly, though, so Ryan and I decided to post an incomplete version of our tech report, and I went back to Mountain View with a vow to go back and fix it once my internship was over.

In August, back at Mozilla, I worked furiously to get at least some part of my traits proposal implemented before my internship ended. I had to give my final presentation on traits in my second-to-last week at Mozilla, before I had anything actually working yet; in it, I overoptimistically predicted that traits would be done by the 0.4 release. (Much of the work ended up being deferred to the 0.5 release, which just came out in December, and many bugs still remain to be addressed.)

But finally, on Monday of my last week, I successfully compiled and ran the first-ever Rust program with a Haskell-style default method. It was just a toy program, but I was pretty excited. I also felt heartened by all the enthusiasm I was seeing for traits, both on the team and in the Rust community beyond the core team. (As I said in my presentation, "I'm an academic -- I'm not used to working on things that people actually want.") For that matter, it was rather exciting that a Rust community beyond the core team could now actually be said to exist; that hadn't been the case back in 2011!

A few days before we had to head back to Indiana, I went to give my talk at Berkeley, where I had a great time chatting with Neil Conway, Joe Hellerstein and the rest of the Bloom and Daedalus folks. Their work on BloomL, which is an extension of Bloom, had turned out to be directly related to our work: what Ryan and I were doing for single-assignment parallel languages -- namely, parameterizing them by a user-specified lattice to give them a more flexible and general notion of monotonicity -- BloomL had already done for Bloom. Their goal was guaranteeing eventual consistency in distributed databases, rather than deterministic parallelism, but the math was largely the same.

Moreover -- and I might never have realized this, had we not come across the Bloom work -- the ideas about monotonicity that Ryan and I had been kicking around had already been in play in the distributed systems community for a while. The novel thing that I hope to be able to offer is a language-based account of those ideas, as part of a language-based approach to deterministic parallelism. But just because something looks like a language problem to me doesn't mean it's okay to succumb to PL myopia! We need to pay attention to what the distributed systems people are doing; it turns out they're working on really hard, interesting problems.

With our internships done, Alex and I flew back home to Indiana, again with cats in tow, and I spent a much-needed week doing absolutely nothing while Alex began teaching NLP to undergrads. On August 27th, we happily celebrated our first wedding anniversary.

Inside "Your rainbow panorama" by Ólafur Elíasson on the roof of ARoS Aarhus Kunstmuseum, Aarhus, Denmark.

The highlight of my September was a trip to Denmark, where I attended ICFP in Copenhagen, then spent a few days adventuring with Chris. While ICFP was going on, Ryan and I received the reviews for our POPL submission and wrote our response to the reviews. I had been nervous that the reviewers would spot the remaining hole in our proof, but my fear turned out to be unfounded -- none of the reviewers took issue with the technical development. Rather, their complaints were that we needed to motivate the paper better and provide more real-world examples of why a new approach to deterministic parallelism was called for, which were both fair points.

Considering that it was my first time submitting a paper to POPL, I was actually quite happy with the outcome: the reviewers found our idea to be interesting and potentially useful, pending more examples. But since it looked like our paper wasn't going to be accepted to POPL, we started thinking about resubmitting it to ESOP. I also began talking to people about potential applications for our work, and Matt Might suggested using LVars for parallelizing CFAs, so I began thinking a bit about how that might work. (I didn't get particularly far with the idea, but as it turns out, two of my favorite colleagues, Will Byrd and Andy Keep, are just about to join Matt's research group as postdocs, and there's a possibility that we'll be collaborating on it in the future, which would be very cool.)

At the end of September, Alex and I went to Florida for a few days to be part of his sister Natalie's wedding, and happily welcomed her new husband, Bryan, into our family.

On October first, Ryan and I were notified that our paper didn't get into POPL, so we spent the first half of the month readying our ESOP submission. Since the ESOP page limit is shorter, we had to condense and cut a lot of the material, which was difficult, but I think ultimately improved the paper. Aaron Turon helped by commenting extensively on a draft. It was also during this time that I finally fixed the renaming bug, after some conversations with Amr and a few emails exchanged with Neel Krishnaswami (who had offered to help when we'd talked at ICFP). We posted an updated version of our tech report with the finished proof.

With the determinism proof finally complete, I went back to thinking about the frame-rule-like property we'd had to show along the way. Something about it was nagging at me. In my experience, people used frame rules to show that, if a program only used certain resources when it ran, then the resources it didn't use remained the same before and after the run. That was a useful property, but it seemed like only half the story. A frame rule could also show that if a program only used certain resources and produced a given result, then running it again with additional resources present would produce the same result.

For us, it was that second guarantee that was particularly important, because "the same result" was what determinism was all about! When I asked about it on cstheory.stackexchange, both Aaron and Neel had illuminating responses, and I realized that there was more to the frame rule than I'd thought. Conveniently, Larry Moss had just asked me to give a talk at the IU logic seminar in November, so I used the classic "agree to give a talk about something as a way of motivating yourself to learn about it" trick and told Larry that I would give a talk about frame properties, adding a paper that Aaron had recommended to my reading list.

Alex all up ons.

Once we'd sent our paper off to ESOP, Ryan suggested that while we were waiting to hear back, I could help out with some monad-par hacking. He asked me to do what he thought of as a fairly minor task, but when I tried to dig into it, I was utterly lost. Since I had barely written any Haskell in my life up to that point, I decided to back up and work on actually learning the language first, which was a bit embarrassing, since Ryan had assumed that I was already pretty fluent with Haskell.

By this point in the year, I was so far behind on my 1000-mile running goal that I more or less gave up on it entirely, although I continued to accrue points on Fitocracy for my bike commutes to campus. Alex, though, continued to run regularly as preparation for another marathon he had coming up in November (he actually did a total of three marathons in 2012). He also became enamored of these things on a weekend trip we made to Atlanta to attend yet another wedding.

I crossed one minor academic cheevo off my list in October: the BloomL paper by Neil Conway et al. was published at SOCC, marking the first time that my work on LVars has been cited by someone else. Finally, in October Larry Moss and Ken Shan both agreed to join Ryan and Amr as members of my dissertation committee, bringing it up to the required total of four members.

As November began, my efforts to get better at Haskell limped along. I had frustrating interactions where I would say that I was struggling with Haskell, and then another grad student would proceed to start 'splaining basic type theory to me. The types weren't the hard part; the purity, the laziness, and the parallelism were the hard part! Eventually, I begged off of working on the monad-par project so that I could get back to my own stuff. I also spent some time organizing events for CS Club, especially the second annual CS Club "So you want to go to grad school?" panel.

Over our Thanksgiving break, Alex and I made a road trip to Iowa to visit my parents and my grandpa Joe, my mom's dad. We spent a few days traipsing around Cedar Rapids, where I lived for a year growing up, and Grinnell, where I went to undergrad.

After Thanksgiving, I gave my logic seminar talk about separation logic and frame properties. While preparing for it, I chatted with Aaron a bit more, and he suggested that I take a look at a paper on "views" that was about to come out at POPL; he suspected there might be an interesting relationship between views and LVars. I began doing some more background reading on concurrent separation logic so that I'd be better able to make sense of the views paper.

Finally, at the end of the month, we got the reviews back from our ESOP submission and wrote our response. The reviewers had liked many things about the paper, but, just as with POPL, they wanted to see more convincing examples of what LVars were good for.

In December, while looking through some old emails, I realized that back in October, Neel Krishnaswami had independently recommended that I read the same paper on views that Aaron had pointed me to. (Dun-dun-DUNNNNNNNN.) I emailed Aaron and Neel to see if they were interested in working with me to investigate a connection between views and LVars, and they responded enthusiastically, inviting me to visit Saarbruecken right after POPL so that we could spend some time working together. There are a couple of other LVar-semantics ideas we're interested in pursuing, as well.

Needless to say, I'm incredibly excited about collaborating with Aaron and Neel. For one thing, I've been wanting to do something involving logical relations and LVars right from the start; for another, the fact that a frame property just fell out in the process of doing the LVar determinism proof suggests the possibility of some kind of separation logic for deterministic parallelism. (Also, I've never been to Germany, and I'm quite looking forward to visiting, especially since Ezgi Cicek alerted me to the existence of Flammkuchen.)

What Will Byrd's office looked like in December 2012, at the end of his nine years at IU.

In mid-December, Ryan and I got a very nice rejection letter from the ESOP committee, in which they reiterated that they liked our idea but that they wanted more convincing examples. Although I'm of course disappointed that we haven't been able to publish our work yet, I think that the two rounds of reviewing we've been through have put us in a good place to move forward with it. The technical development is now quite solid, the idea is novel as far as we know, and people seem to want to be enthusiastic about it. But it's clear that we need to motivate the work better. Everyone knows that nondeterminism can be nasty, but why would someone want to use our thing instead of just using a single-assignment language? What kinds of programs are expressible (or conveniently expressed) with a monotonic-multiple-assignment language that would be impossible (or inconvenient) to express with a single-assignment language or some other existing deterministic parallel model? Those are the questions I think we need to address next time.

The semester wrapped up, concluding with Andy Keep's dissertation defense, and Alex and I skipped town for the holidays, going first to New York for an end-of-the-world party that some friends were throwing, and then to Tallahassee to spend a few days with his family, as has become traditional. All this was great fun, except that just as we got to NYC, I caught a terrible cold that destroyed my voice and just wouldn't go away. When we got back to Bloomington, Alex convinced me to see a professional, who prescribed four different medications. Believe it or not, this is the first time in my adult life that I remember having to take prescription drugs of any kind, and I was a little reluctant to do so, but they seem to have done the job. Still, it's only now, after nine days of drugs, that my throat is getting back to normal.

Finally, I submitted a talk proposal about the LVars work for the POPL student short talk session, and it was accepted, so it looks like I'll be giving a short presentation at POPL in addition to the, uh, even shorter one that I might be giving at PLMW. I have a bit of work to do to prepare for both of those before I head off to POPL (in Rome!) in a couple of weeks.

To sum up, I'm happy with how my 2012 went. I made a lot of progress on research this year. Getting our grant proposal accepted was a big win. I'm proud of having submitted my first paper to POPL, having finally gotten the LVar determinism proof done, and having given various talks and successfully reached out to various people about research collaborations. At Mozilla, I contributed to every release of Rust, and I helped bring traits into the language. In my personal life, my marriage continued to go wonderfully. Alex is still the best. The paper rejections were unfortunate, but that's part of academic life; there's not much I can really complain about.

In 2013, I'd like to do another marathon; I'm tempted to go back to Toronto and try that course again in May. Research-milestone-wise, I want to get our first LVar paper published and get another paper or two underway, and I'm hoping to propose my thesis relatively soon -- sometime this summer, perhaps. Alex is planning to propose this year, too, so he and I are even more or less synchronized, which is awesome. Although we decided we don't want to do any more internships before we graduate, I want to keep contributing to Rust as a volunteer when I can, and to do what I can to help other people contribute. And that's about it for me; thanks for reading! How about you?


[User Picture]From: graydon
2013-01-07 07:51 am (UTC)
It was, as always, great to have your help with rustc again. Pretty soon it'll be all growed up.

Best of luck with the LVars. We need better parallelism models!
(Reply) (Thread)
[User Picture]From: lindseykuper
2013-01-07 07:19 pm (UTC)
Thanks! And thanks!
(Reply) (Parent) (Thread)
[User Picture]From: bubblingbeebles
2013-01-07 06:22 pm (UTC)
i'm surprised by your ability to remember the year at monthly resolution. looking back over my own i see about 6 logical chunks (one of which is "the entire fall semester"... uh oh).
(Reply) (Thread)
[User Picture]From: lindseykuper
2013-01-07 06:36 pm (UTC)
Yeah, months are a pretty arbitrary way of breaking up time, but they have the advantage of being universal. I wrote this entry with a lot of help from email archives, calendar entries, commit logs, and my other journal entries.

And my recollections for the second half of the year are considerably longer than those for the first. I think that's first of all because it happened more recently, so I remember it better, but also because I've had less time to process and mentally summarize it. My brain's good at summarizing, given lots of time.

Edited at 2013-01-07 06:47 pm (UTC)
(Reply) (Parent) (Thread)