Thursday, November 07, 2013

A Theorist Goes to SOSP

Monday I attended the 24th Symposium on Operating Systems Principles, the lead conference for computer systems research. Why would a nice theorist go to SOSP? Trying to recruit a few good systems faculty for Georgia Tech.

I really enjoyed the day in ways I didn't expect. I found several of the talks interesting, even from a theory perspective. Austin Clements, in the first and one of the best paper talks, said he had a theorem and proof (roughly if operations scale there is an implementation that scales well on multicores), though purposely left the formalization and proof out of the talk and focused on implementations. Kay Ousterhout built on some theoretical tools for job scheduling. In a talk after I left, a group from Texas takes a step towards practical proof-based verifiable computing. I never expected to be cited in a SOSP paper.

When I go to a theory conference I see so many people I know that I don't spend enough time meeting new people. At SOSP, I knew a handful of people and just had a great time talking to people I haven't met before, particularly students.

Only thirty papers get presented in single track in this conference held every two years. STOC/FOCS accepts over 300 papers in the same time period. Having an SOSP paper is a really big deal. Despite having only thirty talks and traditionally held in hard-to-reach places (this year an hour and a half drive from Pittsburgh), there were 628 attendees split 42% students, 42% non-student academics, 15% industry and one member of the press.

The 2013 SOSP is the first ACM conference will fully open proceedings and the authors retained full rights to their paper, the gold standard espoused by many in our community. It didn't come cheap, the conference put up $1100/paper to the ACM to pay for the privilege.


  1. Interesting post -- thank you. I'm sorry, but I can't resist the troll bait: what exactly is the service that ACM offered in exchange for $1100/paper? Could this service not have been offered much more cheaply by or some other competitor?

    1. there is more trust that ACM will preserve the papers in the very long run?

  2. Sigh. As someone who has gone to systems-y conferences (though I've never been to SOSP, I've been to SIGCOMM and some others) I find your enjoying yourself unsurprising. They generally have very good papers and talks. There's generally lots of good theory work at systems conferences -- though it wouldn't get the attention of most of the theory community. Their conferences get high attendance so there's plenty of people to talk to. They're also fun people.

    The "practical proof-based verifiable computing" paper you missed is excellent work by a group that's been working in parallel (and, more recently, in conjunction) with (my now-graduated student) Justin Thaler, which naturally has a lot of theory behind it, and the challenges (both theoretical and practical) are how to push the theory to a desirable system.

    I'd expect there to be more collaborations of that type between systems and theory -- and certainly there are some going on all the time -- but for the most part the two communities don't communicate and pay attention to each other. (Actually, I think the systems people are better at paying attention to us than the other way around; I think they know a good theoretical attack for a real problem will mean a good publication, while on the theory side, coming up with an algorithm for a real problem is not necessarily helpful and sometimes outright harmful to your publication chances at a theory conference.)

    Constructively, perhaps the NSF should set up a grant specifically to fund theorists going to systems conferences for the first time and systems people going to theory conferences for the first time. That seems like a possibly effective and relatively cheap way to get the communities to communicate more. (Yes, at most universities, ostensibly a theorist could walk down the hall to talk to a systems person, or vice versa, but generally that doesn't actually happen.)