Hyder: A Transactional Indexed Record Manager for Shared Flash Storage
                Philip A. Bernstein, Microsoft Research

An enormous increase in the I/O rate to shared storage is made
possible by the availability of large flash storage chips and cheap
high-speed network switches. Hyder is a research project to develop a
new transactional indexed-record manager based on these technologies.
It's a data-sharing system, where all compute servers have direct
access to shared flash storage and no direct-attached disk. Its main
feature is that it scales out without partitioning the database or
application. It is therefore well-suited to a data center environment,
where scale-out is especially important and where specialized flash
hardware and networking can be cost-effective. The software
architecture that makes this possible is radically different than
classical transactional record managers. It uses log-structured record
storage, binary search trees, and optimistic concurrency control.
There is no locking, ARIES-style logging, or B-trees. After a brief
discussion of motivation, I will spend most of the talk describing the
architecture. This work is joint with Colin Reid, also at Microsoft.