Follup comments from Jim Larus (Microsoft):

Thanks very much for the invitation to the RAMCloud design review. I enjoyed hearing more about the project and found the discussion with the audience very enlightening.

As promised, I have a few suggestions:

• Look at HW support for things that are hard, in particular the networking and recovery. Raw flash and low-latency interconnects are interesting technologies that are here now. You should find a way to use them, rather than (or in addition to) software solutions.

• Simplify API implementation. The descriptions of the implementations of the various APIs were complex for such simple interfaces. There seemed to be a lot of premature optimization. I assume the system will become simpler as they are better understood.

• Build real apps to drive development. You need use cases and real requirements to drive the system you are building. Without these, you can make up the requirements (1 sec, 10 sec recovery) and declare success.

• You might be interested in our big data store for real-time data: DataGarage: Warehousing Massive Amounts of Performance Data on Commodity Servers

• You might be interested in finding out more about what is going on with real HW these days. In particular, the memory controllers aren't as simple (or fair) as you might assume. This stuff will matter to you.

• Thomas Moscibroda and Onur Mutlu, Memory Performance Attacks: Denial of Memory Service in Multi-Core Systems, Proceedings of the 16th USENIX Security Symposium (USENIX SECURITY), pages 257-274, Boston, MA, August 2007. Slides (ppt)

Thanks again for the invitation, it was a lot of fun,

/Jim