SEDCL Retreat 2014 - Industrial Feedback Session

Pratab Subrahmanyam - VMWare

More information about how to use and setup RAMCloud, would like to use it but it seems information is not public.
Would like RAMCloud to be more open so they can use it.

Ulfar Erlingsson - Google

Missed the first day of talks on RAMCloud overview.
“Two Plus One” and “RAFT” talks were great… represent system ideas that are very interesting as a systems person, and are potentially very useful for all engineers that are coming into companies like Google and starting to build these systems.
Paxos is a black box that’s incredibly complicated, and de-mystifying that is very helpful…. Lamport wasn’t very helpful with his “parliament” ideas.
RAFT is really useful for your average engineer.
Would be great to see more applications with RAMClod and demonstrating the impact of low latency.

Hitoshi Takagi - NEC Japan

Great to educate myself on the group and what’s going on.
As platform vendors for NEC we are looking for eeal solutions for our customers (that’s why we brought the atom machine)
We would like to see real applications with RAMCloud…

Kurt Kolovson - VMWare

Huge fan of this group, highest regard for faculty and graduate students, many thanks for holding these retreats.
Overall very impressed with the quality of these talk…. been two years since last attended, can see a lot of progress has been made since then.
Especially liked the talk by Adam Belay on IX, very worthwhile research project… and also Ankita, liked to see how she has progressed in her research… and also like Henry’s work.
Also liked talk on linearizable RPCs, had a good discussion on linearizability with John, have a personal interest in that. Like to see a view that’s other than the strictly databases point of view.
Talks on “Two Plus One" and “RAFT" and quite relevant to us at VMware, and also the talk on Libra.
Very much enjoyed all the talks.
Nothing but praise for everyone.

Rean Griffith - VMWare

Very much enjoyed the retreat, great to talk to all the students and meet everyone.
Very much enjoyed the SLIK talk, also really liked the latency analysis…. liked to see how you might change the transport under the covers (?)
Besides applications, would also like to see would like to see maybe if you could use RAMCloud to speedup other kinds of queries that might have to do with machine learning applications.
Hopefully can view the slides after the retreat.
Also really enjoyed Asaf’s talk.
Congratulations on getting RAMCloud out there.
Liked the sergeant work, looking forward to more things there.

Daniel Peek - Facebook

Thought the quality of talks was extraordinarily high.
Recommend that for future iterations, we want to see more talks from students and NO talks from industry!
Facebook has been experienced with building systems to solve a specific problem and also without a starting problem.
Those systems that were built without a specific problem in mind tend to burn out very quickly.
Hard to understand what is useful and what is not useful without the applications up front…. advise looking at applications earlier in the development cycle.

Rajesh Nishtala - Facebook

Really liked the talks, but @professors: The scopes of the projects seemed to be much larger than the talks made them out to be
Example: Talk on network partitioning. Understanding the broader scope of this work would have been useful…
Same thing for RAMCloud, latency is great, but how does this apply to Hadoop and other systems.

Naoki Shiota - NEC Japan

One great thing is that RAMCloud has been released and NEC is now creating applications on top of RAMCloud.
Would like to see more applications on top of RAMCloud.
More collaboration between industry and academia.

David Tennenhouse - VMWare

Quality of talks was very good.
People did a good job at explaining the work that they were doing.
Really happy to see a lot of measurement.
Emphasis on latency this time was a lot crisper and clearer. Much more “laser focused”… get’s the differentiator out there much more clearly.
Think it would be a good idea to keep pushing on the tail latency….
If you can really design for microsecond, then how is the world different?
I’m not too sure and I want to be careful with this, but with the secondary indices are those the semantics that people are going to want?
Doing a key value store API is straight forward because people generally agree on what that’s supposed to look like … but there is no such widely agreed upon API for secondary indices in a key value store. Why did you decide on this particular interface?
Really like the talk on IX, thought the talk was very clear.
RAFT: Regarding the publication issue, I’ve noticed that key advances usually appear at USENIX… i.e. kerberos
Who are the students who scored well in the Paxos and RAFT tests? We want to hire them...
@Daniel: Enjoyed the talks from industry… maybe talks about what they would do with a system like RAMCloud.
Would like more poster time.
Did a good job of managing time… but would like to see more poster time.

Dimitri Perelman - Facebook

From our perspective we wanted to say that the design choices that you made are more important that the final product, would like to hear more about that.
Very pleased with talk on IX.
Measurement work in the RAMCloud system is quite interesting because of its high performance. How to do measurement well in a low latency system like RAMCloud is in itself an interesting problem. We lack available tools and techniques today to do this well.

Yuichi Nakamura - NEC Japan

Naoki Nishi - NEC Japan

Two comments:
1) NEC members really hope to find interesting applications for RAMCloud on the new ATOM server.
2) In Japan, we are studying next generation network platform… many telecom operators do not need dedicated hardware appliances for realizing the next generation networks…. they are working on putting network functions on general servers. I feel that some of the things discussed in these talks may also be useful for these next generation networks. After I go back to Japan I would like to consider this issue.

Greg Bollella - VMWare

Lots of praise.
Tail latencies are very important to look at, would like to see more on that.
Thinking about the broader context of the technologies being considered here. There may be some parts of the RAMCloud system that could be used in other systems too.

Jon Affeld - Emulex

Generally datacenter customers do not really have a passion for latency so its interesting to see that here.
Overall quite excited, interesting to see the intelligence at the edge to help bypass the OS, things are getting more powerful for supporting low latency and thus supporting a larger market.

Mark Karnoswki - Emulex

Nice break from the rat race.
Echo comment on industry application… would like to see more applications for RAMCloud.
Challenging the presenters to expand on the ideas that they are presenting to see how it would be useful in other contexts.
Poster session was quite good and I enjoyed talking to people one on one.
My role is to specify our next generation controllers…. so would like to see what is the wish list for what we would like to build in the future or use RAMCloud for.