Abstract. Self-timed scheduling is an attractive implementation style for multiprocessor DSP systems due to its ability to exploit predictability in application behavior, its avoidance of over-constrained synchronization, and its simplified clocking requirements. However, analysis and optimization of selftimed systems under real-time constraints is challenging due to the complex, irregular dynamics of selftimed operation. In this paper, we review a number of high-level intermediate representations for compiling dataflow programs onto self-timed DSP platforms, including representations for modeling the placement of interprocessor communication (IPC) operations; separating synchronization from data transfer during IPC; modeling and optimizing linear orderings of communication operations; performing accurate design space exploration under communication resource contention; and exploring alternative processor assignments during the synthesis process. We review the structure of these representations, and discuss efficient techniques that operate on them to streamline scheduling, communication synthesis, and power management of multiprocessor DSP implementations.

Abstract-Multiprocessor mapping and scheduling algorithms have been extensively studied over the past few decades and have been tackled from different perspectives. In the late 1980's, the two-step decomposition of scheduling-into clustering and clusterscheduling-was introduced. Ever since, several clustering and merging algorithms have been proposed and individually reported to be efficient. However, it is not clear how effective they are and how well they compare against single-step scheduling algorithms or other multistep algorithms. In this paper, we explore the effectiveness of the two-phase decomposition of scheduling and describe efficient and novel techniques that aggressively streamline interprocessor communications and can be tuned to exploit the significantly longer compilation time that is available to embedded system designers. We evaluate a number of leading clustering and merging algorithms using a set of benchmarks with diverse structures. We present an experimental setup for comparing the single-step against the two-step scheduling approach. We determine the importance of different steps in scheduling and the effect of different steps on overall schedule performance and show that the decomposition of the scheduling process indeed improves the overall performance. We also show that the quality of the solutions depends on the quality of the clusters generated in the clustering step. Based on the results, we also discuss why the parallel time metric in the clustering step may not provide an accurate measure for the final performance of cluster-scheduling.

In this paper, we present a modular co-synthesis framework called CHARMED that solves the problem of hardware-software co-synthesis of periodic, multi-mode, distributed, embedded systems. In this framework we perform the synthesis under several constraints while optimizing for a set of objectives. We allow the designer to fully control the performance evaluation process, constraint parameters, and optimization goals. Once the synthesis is performed, we provide the designer a non-dominated set (Pareto front) of implementations on streamlined architectures that are in general heterogeneous and distributed. We also employ two different techniques, namely clustering and parallelization, to reduce the complexity of the solution space and expedite the search. The experimental results demonstrate the effectiveness of the CHARMED framework in computing efficient co-synthesis solutions within a reasonable amount of time.

For multiprocessor embedded systems, the dynamic voltage scaling (DVS) technique can be applied to scheduled applications for energy reduction. DVS utilizes slack in the schedule to slow down processes and save energy. Therefore, it is generally believed that the maximal energy saving is achieved on a schedule with the minimum makespan (maximal slack). Most current approaches treat task assignment, scheduling, and DVS separately. In this paper, we present a framework called CASPER (Combined Assignment, Scheduling, and PowER-management) that challenges this common belief by integrating task scheduling and DVS under a single iterative optimization loop via genetic algorithm. We have conducted extensive experiments to validate the energy efficiency of CASPER. For homogeneous multiprocessor systems (in which all processors are of the same type), we consider a recently proposed slack distribution algorithm (PDP-SPM) [3]: applying PDP-SPM on the schedule with the minimal makespan gives an average of 53.8% energy saving; CASPER finds schedules with slightly larger makespan but a 57.3% energy saving, a 7.8% improvement. For heterogeneous systems, we consider the power variation DVS (PV-DVS) algorithm [13], CASPER improves its energy efficiency by 8.2%. Finally, our results also show that the proposed single loop CASPER framework saves 23.3% more energy over GMA+EE-GLSA [12], the only other known integrated approach with a nested loop that combines scheduling and power management in the inner loop but leaves assignment in the outer loop.

In this paper, we address two key trends in the synthesis of implementations for embedded multiprocessors-(1) the increasing importance of managing interprocessor communication (IPC) in an efficient manner, and (2) the acceptance of significantly longer compilation time by embedded system designers. The former aspect is especially evident in the increasing interest among embedded system architects in innovative communication architectures, such as those involving optical interconnection technologies, and hybrid electro-optical structures [8, 19]. The latter aspect-increased compile-time tolerance-results because embedded multiprocessor systems are typically designed as final implementations for dedicated functions. While multiprocessor mapping strategies for general-purpose systems are usually designed with low to moderate complexity as a constraint, embedded system design tools are allowed to employ more thorough and time-consuming optimization techniques [13].