All computations are created equal, but some are more equal than others: Novel designs and uses of communication in auxiliary processing systems.

摘要

As clock speeds and transistor densities continue to increase, cheap, fast, parallel single-chip systems are being introduced in the commodity market. Placing high-performance, homogeneous parallel processing power on-chip has the economic advantage of designing a single component and replicating it several times on the chip, saving design and verification time. There is the potential, however, for heterogeneous designs as well. Computation is not always the bottleneck for performance—memory bandwidth and latency are. Adding computation power on-chip does helps little relieve memory pressure. In addition, parallelization of existing programs is not trivial. To increase single-thread performance in commodity systems, high-performance parallelism must be achievable by the average programmer.; In this thesis, I examine three heterogeneous “auxiliary” processor systems. All three systems have the same goal—parallelize a computation to better utilize the available resources—but the routes taken are very different. The first, Active Pages, places processing in memory to allow computations to take advantage of high on-chip bandwidth and low latencies. This is designed for a sophisticated programmer to get the most out of a new architecture. The second, a software security system, leverages a very small amount of chip space to dramatically increase the performance of dynamic pointer checking. This parallelism is provided entirely by the system, and is transparent to the programmer, making a major high-overhead debugging task more efficient. Finally, I present IOP (Instruction-level Object Parallelism), a system that provides an instruction-level interface to exploit object-level parallelism. This leverages the practices already used in object-oriented programming to expose parallelism with the low overhead communication and synchronization in modern processors, providing semi-transparent, high-performance parallelization.; In all three systems, communication mechanisms are compared based on hardware costs, performance, and programmability. Because the target is commodity desktop systems, existing communication mechanisms will be used as often as possible.