Calendar

Hot-Potato I/O: Improving and simplifying I/O DMA transactions on cluster systems
Performance improvement of computer system I/O has been slower than processor and memory technologies in terms of reducing latency, increasing bandwidth, and other factors. Based on this observation, how I/O is performed needs to be re-examined and explored for optimizations. To rebalance the computer system having multiple cores and multiple CPUs with integrated memory controllers, this paper proposes a new system oriented I/O method where data movement is controlled more directly by the CPU. This is achieved by treating the ownership of payload data movement as a “Hot-Potato” and the I/O interfaces as simple FIFOs. Our experiments with the proposed method on a current system show that transmit latency for small Ethernet frames reduces from 1,673 ns to 108 ns, and memcached application testing shows 8% latency reduction. Several less quantifiable benefits that exist in terms of system power management, I/O quality-of-service, and system complexity are also discussed.

Major Advisor: Ben Lee
Committee: Bella Bose
Committee: Thinh Nguyen
Committee: Bechir Hamdaoui
GCR: Guenter Schneider 

• Electrical Engineering and Computer Science*