The seminar will meet Mondays 10-12am at Davis 113A. The prerequisite is that the student must have completed CSE 586: Distributed Systems or equivalent graduate level distributed systems course. Here is a link to the seminar information.

We will discuss 14 papers in 14 weeks. Each student will serve as a presenter for one paper, and as a participant for all the remaining papers. We will run the seminar as follows: (Here is another description of the format that explains the motivation behind it.)

1) By Saturday midnight, each participant contributes 1 or 2 questions about the paper to the course webpage at Piazza.com.

2) The presenter writes a 2-3 page review of the paper, prints the copies of his review, and brings this to the meeting for review before his presentation. All the participants review the hardcopy of the review-report in the first 15 minutes of the meeting. So in the first 15 minutes, it is all silence. The presenter is at the podium, getting adjusted, while we review his report.

3) The presenter will use 40 minutes to discuss the heart (the most important and useful part) of the paper. The presenter is allowed to use no more than a 10-15 slide presentation. The slides should be mostly for visuals: figures, tables, graphs.

4) After the presentation, the presenter starts answering questions with first those that were submitted 1-2 day in advance on Piazza. The audience members who submitted the questions read the questions loudly, the presenter answers. And then we have more questions from the floor and more comments from the floor.

5) After the question-answer phase, participants form groups of 3, and mock-review the paper. The groups write the review collaboratively using Google docs. In a group, one participant may research about related work, and write that part of the review, while another may write about motivation/application aspect of the paper, and the other about the technical/methods aspects. Instead of a review a group can also write about related research questions to the current paper, in order to come up with interesting (and secondarily actionable) directions for future work. Again the group needs to be aggressive in its effort and brainstorm to come up with "novel", albeit speculative/little-far-fetched, research ideas.

- Large Scale Distributed Deep Networks (NIPS 12)
- TensorFlow: A System for Large-Scale Machine Learning (OSDI 16)
- Incremental, iterative data processing with timely dataflow (COMM ACM 16)
- Prioritizing Attention in Fast Data: Principles and Promise (CIDR 17)
- Large-scale cluster management at Google with Borg (Eurosys 15)
- Slicer: Auto-Sharding for Datacenter Applications (OSDI 16)
- Musketeer: all for one, one for all in data processing systems (EuroSys 15)
- Making Sense of Performance in Data Analytics Frameworks (NSDI 15)
- The Bitcoin Backbone Protocol: Analysis and Applications (2015)
- The many faces of consistency (2016)
- The SNOW Theorem and Latency-Optimal Read-Only Transactions (OSDI 16)
- Incremental Consistency Guarantees for Replicated Objects (OSDI 16)
- Just Say NO to Paxos Overhead: Replacing Consensus with Network Ordering (OSDI 16)
- FaSST: Fast, Scalable and Simple Distributed Transactions with Two-Sided (RDMA) Datagram RPCs (OSDI 16)

As a participant, you should expect to spend around 4 hours reading each paper. When it is your turn to become the presenter, the preparation time should take to the north of 16 hours. Go through a couple papers in the list to gauge if you can handle the material technically.

Useful linksHow I read a research paper

How to present your work

How to write your research paper

How I write

Advice to graduate students

Each student will be evaluated the same way, regardless of whether she is taking the class for 1 credit or 3 credits. I will assign the S/U grade based on attendance and paper presentation and discussion performance. (As department policy, we do not assign letter grades for 700-level seminar courses.)

Link to Spring 2016 seminar and reading list.