Abstract

Commodity disk IO architectures suffer the weaknesses of (i) lowered performance from kernel-user memory copies, (ii) poor cross-process isolation, (iii) poor resource management and accounting at kernel level, and (iv) inability for processes to control their data caching, transfer scheduling and block placement.

Container IO (CIO) is a new IO architecture designed to address the described weaknesses by providing access to filesystems at the logical block level. Processes operate under the notion of exclusive access to the filesystem, however CIO is instrumental in updating filesystem state, providing concurrency support, enforcing access control and providing processes with data sharing and transfer primitives.

In this talket I (tersely) describe the architecture; specifically the abstractions defined, their interaction, constraints on filesystem design and the issues arising from providing processes direct control over data caching and transfer scheduling. Finally, I present some performance results from a prototype implementation.