EVALUATING THE LOCALITY BENEFITS OF ACTIVE MESSAGES

A major challenge in fine-grained computing is achieving locality with out excessive scheduling overhead. We built two J-Machine implementati ons of a fine-grained programming model, the Berkeley Threaded Abstrac t Machine. One implementation takes an Active Messages approach, maint aining a scheduling hierarchy in software in order to improve data cac he performance. Another approach relies on the J-Machine's message que ues and fast task switch, lowering the control costs at the expense of data locality. Our analysis measures the costs and benefits of each a pproach, for a variety of programs and cache configurations. The Activ e Messages implementation is strongest when miss penalties are high an d for the finest-grained programs. The hardware-buffered implementatio n is strongest in direct-mapped caches, where it achieves substantiall y better instruction cache performance.