Compiler Technology for Scalable Servers

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

This research investigates compiler techniques for improving the performance and scalability of Internet server programs. Servers have a different set of characteristics from the computations that compiler writers have focused on in the past. Instead of executing a compute-bound computation with a modest number of long-lived threads, servers typically use many short-lived threads to manage I/O bound connections from multiple clients. These new characteristics place a premium on previously less relevant aspects of the system such as the thread creation overhead and thread stack overhead. They also provide opportunities that the compiler can exploit to improve the memory management. The research will attack thread overhead by automatically transforming multithreaded code to event-driven form. Code in event-driven form contains an event loop that repeatedly blocks waiting for the next event from any of the current connections. It then invokes a handler that executes the appropriate action, then returns to the event loop. This transformation will eliminate thread creation and management overhead. The compiler will attack memory management overhead by analyzing the program to identify objects whose lifetimes are tied to individual threads, then using specialized memory management algorithms for these objects.