3rd Workshop on Compiler-Driven Performance

October 06, 2004
Hilton Suites Toronto/Markham Conference Centre
Markham, ON

Associated with CASCON 2004
(http://www.cas.ibm.com/cascon)


 

Accelerating Java synchronization in Just-In-Time compiler-generated code

Mark Stoodley - IBM Toronto Lab

Synchronization is an essential tool that is widely used in Java applications and in the Java class library to maintain data integrity in a multithreaded environment. Locking and unlocking objects are generally expensive operations, however, so Java Just-In-Time(JIT) compilers employ a variety of ingenious tactics to reduce the cost of synchronization. These tactics fall into two broad categories: 1) techniques to reduce the cost of the lock and unlock operations, and 2) techniques to eliminate some of the operations altogether. The use of flat locks and spin loops are two techniques in the first category that reduce cost by avoiding the use of monitors at the operating system level in the typical case that there is no thread contention for an object's lock. Nested lock removal and lock coarsening are examples of techniques in the second category that operate by removing redundant synchronization operations or by combining the effects of multiple operations so that the cost is not paid as often. In this talk, I will first give a quick overview of the various technologies used to improve the cost of synchronization in Java programs. I will then describe in greater detail the lock coarsening technique used in recent releases of IBM's JIT compilers to broaden the scope of code across which locks are held so that fewer lock operations are required. While improving performance is the obviously the main goal of this optimization, maintaining correctness is critically important. I will describe two extremely subtle problems that would have sacrificed correctness for performance had our framework not been specifically designed to detect and avoid them.

Presentation Slides.