ACAL - Research Briefs

Peter M. Chen


Rio: High-Performance File Systems via Reliable Memories

Graduate Students: Wee Teck Ng, Gurushankar Rajamani, Rajagopalan Sivaramakrishnan
Faculty: Peter M. Chen and Trevor Mudge
Sponsor: National Science Foundataion, Digital Equipment Corporation, Rackham Graduate School

The last decade's phenomenal improvement in computer processing power has made current storage technology woefully inadequate. The speed mismatch between processors and magnetic disks has hampered applications such as electronic libraries, scientific visualization, and large-scale simulation. Data-intensive applications now run at disk speeds and cannot benefit from further processor improvements.

Our project, Rio (RAM I/O), seeks to radically improve storage performance by making it practical to store files long-term in memory. To do this, memory must be able to store data as reliably as disks do; this means surviving system crashes.

We have modified the Digital Unix operating system on DEC Alpha workstations to protect memory during system failures and to restore the contents of memory upon reboot. We have shown that our system enables memory to survive 99% of all system crashes, which is comparable to disk reliability levels. These results elevate memory to a first-class citizen of the storage hierarchy and enable many new optimizations for performance and reliability.

Future works includes implementing Rio in Windows NT, measuring memory's resistance to database crashes, and using reliable memory to accelerate distributed applications.