Polymorphous architectures: a unified approach for extracting concurrency of different granularities

Abstract

Processor architects today are faced by two daunting challenges: emerging applications with heterogeneous computation needs and technology limitations of power, wire delay, and process variation. Designing multiple application-specific processors or specialized architectures introduces design complexity, a software programmability problem, and reduces economies of scale. There is a pressing need for design methodologies that can provide support for heterogeneous applications, combat processor complexity, and achieve economies of scale. In this dissertation, we introduce the notion of architectural polymorphism to build such scalable processors that provide support for heterogeneous computation by supporting different granularities of parallelism. Polymorphism configures coarse-grained microarchitecture blocks to provide anadaptive and flexible processor substrate. Technology scalability is achieved by designing an architecture using scalable and modular microarchitecture blocks. We use the dataflow graph as the unifying abstraction layer across three granularities of parallelism-instruction-level, thread-level, and data-level. To first order, this granularity of parallelism is the main difference between different classes of applications. All programs are expressed in terms of dataflow graphs and directly mapped to the hardware, appropriately partitioned as required by the granularity of parallelism. We introduce Explicit Data Graph Execution (EDGE) ISAs, a class of ISAs as an architectural solution for effciently expressing parallelism for building technology scalable architectures. We developed the TRIPS architecture implementing an EDGE ISA using a heavily partitioned and distributed microarchitecture to achieve technology scalability. The two most signicant features of the TRIPS microarchitecture are its heavily partitioned and modular design, and the use of microarchitecture networks for communication across modules. We have also built aprototype TRIPS chip in 130nm ASIC technology composed of two processor

cores and a distributed 1MB Non-Uniform Cache Access Architecture (NUCA) on-chip memory system. Our performance results show that the TRIPS microarchitecture which provides a 16-issue machine with a 1024-entry instruction window can sustain good instruction-level parallelism. On a set of hand-optimized kernels IPCs in the range of 4 to 6 are seen, and on a set of benchmarks with ample data-level parallelism (DLP), compiler generated code produces IPCs in the range of 1 to 4. On the EEMBC and SPEC CPU2000 benchmarks we see IPCs in the range of 0.5 to 2.3. Comparing performance to the Alpha 21264, which is a high performance architecture tuned for ILP, TRIPS is up to 3.4 times better on the hand optimized kernels. However, compiler generated binaries for the DLP, EEMBC, and SPEC CPU2000 benchmarks perform worse on TRIPS compared to an Alpha 21264. With more aggressive compiler optimization we expect the performance of the compiler produced binaries to improve. The polymorphous mechanisms proposed in this dissertation are effective at exploiting thread-level parallelism and data-level parallelism. When executing four threads on a single processor, significantly high levels of processor utilization are seen; IPCs are in the range of 0.7 to 3.9 for an application mix consisting of EEMBC and SPEC CPU2000 workloads. When executing programs with DLP, the polymorphous mechanisms we propose provide harmonic mean speedups of 2.1X across a set of DLP workloads, compared to an execution model of extra ting only ILP. Compared to specialized architectures, these mechanisms provide competitive performance using a single execution substrate.