A case study of a hardware-managed TLB in a multi-tasking environment

There have been very few performance studies of hardware-managed translation look-aside buffers (TLBs).The major reason is the lack of efficient and accurate analysis tools. Newer operating systems, applications, and the popularity of the client-server model of computation place a greater burden than their predecessors on memory system components such as TLBs. Thus it is becoming more important to measure the performance of memory systems under such workloads. In this work, we implemented a trap-driven simulator in an operating system to emulate a variety of TLBs. Using this tool, we were able to evaluate the performance of a range of TLBs under these newer workloads. The results show that in order to improve the TLB performance, we should carefully map pages into the TLB, append process identifiers to avoid flushing the TLB contents frequently, or reserve part of the TLB for a particular server process.