面向对象语言并行化中的调用局部化优化

该文提出了一种将调用局部化技术应用于并行环境下面向对象语言的方法，文中详细讨论了该技术的适用条件以及如何通过该方法减少循环中的远程过程调用开销，该优化技术产首先将循环分离成多个包含有远程调用的循环，再将分离后的循环分离给循环中对象所在的处理器，最后，化简迭代空间，并且用消息传递来传输数据，这种优化对象分布和循环并行化之后进行，将函数调用局部化于处理器，通过这种优化，可以进一步挖掘循环中的任务并行性，降低计算复杂度，减少函数调用开销，尤其适合面向对象语言中对循环里小函数的优化，该技术已经在作者设计的Java自动并行化编译器JAPS－Ⅱ中实现，在实验中，利用这种优化技术得到了超线性性加速比。

Method Invocation Localizing Optimization in Parallelizing Object-Oriented Language

Concurrent management and data distribution problems in parallel program design frustrate the construction of large scale software. In recent years, the researches to combine both parallel processing and object oriented techniques have got many result, including explicit parallel languages and automatic parallelizing compilers. Object oriented languages have many features that traditional languages have not. Java even has many new parallel facilities of modern languages. These features greatly frustrate the direct application of traditional techniques of parallelization to object oriented languages. For example, many small functions in object oriented programs are main obstacles to optimize the performance because these small functions generally have little computing cost but quite a bit invocation cost. In this paper, we propose one method for parallelizing OOL to reduce RPC cost. We discuss the condition for applying our techniques and introduce how to apply our techniques to reduce RPC cost. The optimization firstly separates one loop into multiple loops, and then distributes separated loops into different processors where the invoked object in loop is allocated. Finally, we reduce the iteration space and transports data by message passing. We apply the optimization after distributing objects and parallelizing loops. The optimization localizes method invocation at the processor. Through the optimization we exploit further parallelism, reduce computing complexity and reduce function call cost. The optimization is suitable for small function optimization in OOL. We have implemented the optimization in our automatic parallelizing compiler for JAVA, JAPS II. Our experiment demonstrates ultra linear speedup.