Anemone

Memory hungry applications consistently keep their memory requirement curves ahead of the growth of DRAM capacity in modern computer systems. Such applications quickly start paging to swap space on the local disk, which brings down their performance, an old and ongoing battle between the disk and RAM in the memory hierarchy. This thesis presents a practical low-cost solution to this important performance problem. We give the design, implementation and evaluation of Anemone - an Adaptive NEtwork MemOry engiNE. Anemone pools together the memory resources of many machines in a clustered network of computers. It then presents an interface to client machines in order to use the collective memory pool in a virtualized manner, providing potentially unlimited amounts of memory to memory-hungry high-performance applications. Using real applications like the ns-2 simulator, the ray-tracing program POV-ray, and quicksort, disk-based page-fault latencies average 6.5 milliseconds whereas Anemone provides an average of latency of 700.2 microseconds, 9.2 times faster than using the disk. In contrast to the disk-based paging, our results indicate that Anemone reduces the execution time of single memory-bound processes by half. Additionally, Anemone reduces the execution times of multiple, concurrent memory-bound processes by a factor of 10 on the average. Another key advantage of Anemone is that this performance improvement is achieved with no modifications to the client's operating system nor the memory-bound applications due to the use of a novel NFS-based low-latency remote paging mechanism.