Latency to memory is ever-increasing and has reached the point where conventional computer architectures are hitting a performance "wall." Concurrently, several emerging technologies combine processing and memory in a single device. It is expected that in the near future nearly all integrated circuits will have both logic and memory, called "processing in memory." This research creates "collaborative memory" an active memory system that abstractly provides processing in memory. It increases the end-to-end performance of applications and exposes additional parallelism, and mitigates the effect of memory latency. This research builds on the Modify-on-Access (MonA) file system, an active file system that performs operation on behalf of an application, which has been develop under the direction of the PI. The research extends the MonA file system in two ways. One way is closer to the microprocessor, in the virtual memory system, through fine-grain collaborative memory operations visible to an application. The other way is farther from the microprocessor, in an intelligent peripheral device. This offloads computation from the main microprocessor to the peripheral microprocessor. Together the two extensions create a system that increases performance, exposes implicit parallelism, and tolerates memory latency without new advances in compiler technology or programming languages.
