| In the present computing environment, clients are connected by a local area network or a wide area network to servers, which are built from either an individual machine or a cluster of machines. Current technology trends indicate an increasing gap between processor and disk speed, implying that memory needs to be managed more efficiently in order to support high performance networked servers.; This dissertation provides solutions to four important problems in memory management for networked servers: server caching, application data locality, volatile memory and shared memory design tradeoffs . First, this dissertation characterizes the access behavior of server caches, identifies the key properties a good server cache replacement algorithm should have, proposes a new replacement algorithm, and shows that it performs better than all tested alternative algorithms. Second, this dissertation presents a run-time scheduling approach called fine-grained thread scheduling to improve data locality for both sequential and parallel applications. Third, this dissertation introduces a fast, automatic failover approach using efficient memory mirroring in cluster based servers. Finally, this dissertation investigates the tradeoffs between coherence granularities and coherence protocols for software distributed shared memory systems.; The main conclusion of this work is that using innovative methods to manage memory can significantly improve the performance of a fault tolerant networked server. To achieve this requires studying memory system behavior, understanding performance bottlenecks, investigating different approaches, taking advantage of new hardware, and designing new algorithms. |
