James Mickens, Microsoft Research, Distributed System Group
MapReduce: boring...zzzzzzzz
Cloud is dirty underneath...
Unmodified POSIX/Win32 applications running with cloud scale I/O performance
App
--------------------------
Blizzard virtual drive
---------------------------
Disks
1. Delayed durability semantics: flush is an order semantic only, not durability semantic anymore
2. How to avoid buffer a lot of data: treat back storage as a distributed log, and only checkpoint up to the durability point
Now:
All previous writes are durable
All after writes are not durable (which is nice...)
But still, it might be impossible to have the cloud work
Security: government is more powerful than you give it credit for. Using cryptography is kind like using a gun to defend against the government....
Cloud Security:
Even if you encrypt all your data, cloud provider could still monitor your computation (search and fetch, etc.) and derive correlations.
