Michael Walker

ICFP: Retrospective

Looking back at ICFP nearly one month later.
Published on

Almost a month ago now I was heading off to spend a week at ICFP and related events, so here’s a little trip report.

Déjà Fu

My Déjà Fu (git) talk went down pretty well, and there is a recording available. A few people asked questions or talked to me afterwards.

One questioner seemed particularly bothered by the fact that modifyCRef acts like atomicModifyIORef. I’m not sure what he wanted me to say, but that’s ok because Ryan Newton then had very favourable comments, and suggested that Déjà Fu could be the tool around which all the deterministic parallelism libraries (like Par and LVish) could unify around, which is awesome!

(Un?)Fortunately, doing this absolutely requires Déjà Fu to deal with relaxed memory weirdness, as LVish needs compare-and-swap. Fortunately I found a beautiful approach to solving that problem today! Going to implement that next week.

Ryan and I have exchanged a couple of emails as we figure out the best way to integrate the abstractions in Déjà Fu into other libraries, in order to make them testable.

Search Party

The Search Party (git) poster that I submitted for the Student Research Competition didn’t go down so well. People seemed interested that parallelising list comprehensions could speed things up, but it seems that the judges in particular felt that there wasn’t really much novelty.

Unfortunately, there was no feedback given for entries which didn’t pass the poster round.

Presentations I Particularly Liked

  • Partial Aborts for Transactions via First-Class Continuations is about partially undoing transactions in STM, to avoid repeating lots of work. Before the talk, I had no idea how such a thing could be implemented soundly, so it was really enlightening.

  • A Fast Compiler for NetKAT is about a compiler from a network control language using an abstract topology down to concrete programs for the actual switches to run.

  • Functional Pearl: A Smart View on Datatypes is about extending data structures with additional constructors for performing traditionally expensive operations (such as concatenating lists) in constant time, and using smart patterns to rebalance the structure when it is deconstructed through pattern matching, giving better performance overall.

  • Efficient Communication and Collection with Compact Normal Forms is about a way of storing data structures compactly in the heap to facilitate easy network transfer. An interesting approach as it doesn’t implement a lot of common wisdom, and relies on virtual memory.

  • Deja Fu: A Concurrency Testing Library for Haskell this was an amazing talk given by a true genius.

In addition, the contest presentation was really fun.

Other Things

I asked both Simon Peyton Jones and Simon Marlow about research internships. SPJ said he doesn’t do much concurrency, but wouldn’t necessarily rule anything out. Simon Marlow said that he’ll be looking for a research intern at Facebook over the next year, and will let me know when they start taking applications. In my follow-up email to him I mentioned my prior work on garbage collection, as in his talk in PLMW he said he would quite like to return the GHC’s parallel garbage collector at some point.