| key | value |
|---|---|
| When: | Tuesdays and Thursdays at 9:50am |
| Where: | Engineer 2 192 |
| Who: | Peter Alvaro |
| Office (half-)hours: | Tuesdays 11:30-12 (in my office), Fridays 2-2:30pm (zoom) |
| Prerequisites: | A background in systems, an interest in theory, and a sincere curiosity about how stuff works and breaks |
This graduate reading seminar will be dedicated to award-winning papers in systems and data management. We will cover a combination of very recent and classic papers in SOSP, OSDI, NSDI, SIGMOD, and VLDB, with an emphasis on current work.
With very few exceptions, computer scientists send their best work to peer-reviewed conferences that are typically held annually. During these 2-4 day events, papers accepted to the conference are presented by one of the authors. But these talks, which are rarely recorded, are ephemeral. The real artifacts are the papers, which then appear in the ``proceedings'' of the conference.
As many of you already know, getting a paper accepted to a conference is hard. 3-5 anonymous reviewers have to reach consensus that the community will learn something from the paper. Often, they will request significant changes to a paper before reconsidering it in a resubmission, and hence sometimes we find ourselves publishing a paper that we would not have written as it is, in order for it to pass the filter. This is still of course the lucky case -- a majority of papers are invariably rejected, and a low acceptance rate is often considered the mark of a top conference.
Other papers meet a very different fate. As part of conference organization, a paper (sometimes two) is selected as "best paper" for that conference year, and the authors are presented with an award. These awards are extremely prestigious. Although the criteria by which they are selected is clearly subjective, it is no more so than the criteria used to choose papers to appear in the conference! What makes these papers so uncontroversially significant? Are there winning patterns in terms of presentation? Do all best papers resemble one another?
Conferences typically also present awards to past papers that have been highly influential. The criterion for evaluating these "test of time" or "hall of fame" papers is much more objective: papers can be selected based on their citation count and other impact measures, technology transfer to industry, and other criteria. Still, the reasons that these papers became so influential can remain mysterious! What are the hallmarks of a lasting contribution to the discipline? Do all classic papers resemble one another?
This course is a research seminar: we will focus primarily on reading and discussing conference papers. We will read 1-2 papers (typically 1) per session; for each paper, you will provide a brief summary (about 1 page). The summary should answer some or all of the following questions:
- What problem does the paper solve? Is is important?
- How does it solve the problem?
- What alternative approaches are there? Are they adequately discussed in the reading?
- How does this work relate to other research, whether covered in this course or not?
- What specific research questions, if any, does the paper raise for you?
- Why do you think this paper was selected for the best paper award?
The tentative reading list for this class (a work in progress) can be found here:
All students will be expected to present at least two research papers to the class. The presentation format is open: students will not be required to prepare slides (though doing so may help to organize the subject matter). If you do use slides, they should be original (created by you), and external material such as figures should be properly attributed. If you choose not to present slides, use the whiteboard well.
Viewed through one lens, this is the easiest class you are ever going to take. All you need to do is read the papers, show up, and participate in discussion about the papers. Since you are a grad student (most of you, anyhow) part of the reason you are here is studying the state of the art in research -- this class will force you to do this, albeit in a specific area (systems). You are not required to take exams, do homework beyond reading and short summaries, or undertake a final project. There is no writeup! This is JUST a reading seminar, and the reading material is good! Best class.
Viewed through another lens, this is the hardest class you are going to take here. I expect you to read the papers, carefully, to come to class, and to actively participate in discussion. As you will see below, you are evaluated only on this basis. This unfortunately means that even if you do all the readings, write all the summaries, and show up to class every day, you will not pass the class unless you contribute regularly to discussion (which of course requires carefully having read the papers). Reading papers and discussing them is a LOT of work! But it is what we are here to do.
If for some reason you will not be able to participate in discussion, but nevertheless want to take this class, please talk to me in office hours as soon as possible.
| Subject | Share |
|---|---|
| Paper summaries | 20% |
| Participation | 60% |
| Presentations | 20% |
The fact that participation accounts for 60% of the grade should give you an idea of the important of class attendance.
Collaboration is a key part of research. I encourage you to discuss the readings and the final project ideas with your classmates. However, you must reveal the students with whom you discussed the papers in your summaries. Failure to do so will result in formal disciplinary proceedings.
I should not need to say so, but I do: plagiarism in any form is not acceptable and will not be tolerated. As researchers, we always stand upon the shoulders of giants, and building upon existing work is the norm. It is essential, however, that we provide proper attribution. When in doubt, cite! Missing a relevant piece of related work is an embarassment for any researcher; failure to cite a direct influence is dishonest and catastrophic to a researcher's career. The related work section is one of the most critical parts of any paper: spend an adequate amount of time on it.