Performance and security in parallel and distributed systems

点击翻译按钮获取中文摘要

Distributed caching involves storing data in strategic locations within a network and is well-understood in the context of file access, memory access and media content access where it has been studied for many years. In each new environment where caching is implemented, four main questions must be answered: how much information should be stored, where should it be stored, how many copies should be kept at one time, and how does the system determine what is valuable? Depending on system functionality, the type of distribution, system scale, and the amount of parallelism, the answers differ. In this research program, I will investigate several representative emerging systems to explore the costs/benefits of these caching principles, as they apply. Four examples of systems where the replication, location and valuation strategies are expected to differ are 1) malware signatures/traffic profiles in network intrusion detection systems, 2) data and instruction caches in multi-core processors, 3) video objects/metadata in media content distribution networks, and 4) database/web content in multi-tier client-server applications. The expected outcomes are general principles as well as context-specific analysis tools and prototype systems that will give insight to designers in building efficient and secure distributed systems. As a concrete example, systems pay for security with performance, since the extra processing needed to encrypt data or to check for malware is taken from the system's primary tasks. This is a minor issue for network intrusion detection using desktop computers, but is crippling to resource-challenged, large-scale systems, such as wireless mesh and sensor networks that lack resources for all nodes to perform sophisticated traffic filtering. Caching data, malicious code profiles and/or traffic statistics limits exposure of raw data to the insecure network and enables selective traffic filtering that provides high protection power per byte. I believe that ideas from my previous caching experience applied to this new area has the potential to yield novel results.