[Paper Review] A Comparative Study of Scheduling Techniques for Multimedia Applications on SIMD Pipelines

Mehmet Ali Arslan, Flavius Gruian+1 more|arXiv (Cornell University)|Feb 26, 2015

Video Coding and Compression Technologies28 references3 citations

TL;DR

This paper evaluates three scheduling techniques—ad hoc overlapped execution, modulo scheduling, and modulo scheduling with unrolling—for multimedia workloads on SIMD pipelines. It demonstrates that modulo scheduling with unrolling achieves the highest throughput with manageable resource costs, offering a practical trade-off for streaming applications on specialized parallel architectures.

ABSTRACT

Parallel architectures are essential in order to take advantage of the parallelism inherent in streaming applications. One particular branch of these employ hardware SIMD pipelines. In this paper, we analyse several scheduling techniques, namely ad hoc overlapped execution, modulo scheduling and modulo scheduling with unrolling, all of which aim to efficiently utilize the special architecture design. Our investigation focuses on improving throughput while analysing other metrics that are important for streaming applications, such as register pressure, buffer sizes and code size. Through experiments conducted on several media benchmarks, we present and discuss trade-offs involved when selecting any one of these scheduling techniques.

Motivation & Objective

• To evaluate and compare scheduling techniques tailored for SIMD pipelines in multimedia streaming applications.
• To identify optimal scheduling strategies that maximize throughput while managing resource constraints such as register pressure and buffer size.
• To analyze the trade-offs between code size, execution efficiency, and resource utilization across different scheduling methods.
• To provide practical insights for compiler developers and system architects targeting embedded image processing systems.

Proposed method

• Ad hoc overlapped execution is applied by manually scheduling operations to overlap computation and data movement across pipeline stages.
• Modulo scheduling is used to restructure loops to exploit ILP, scheduling operations to minimize initiation interval and maximize pipeline utilization.
• Modulo scheduling with loop unrolling increases ILP by replicating loop bodies, reducing initiation interval and improving throughput.
• The techniques are evaluated using a set of multimedia benchmarks targeting a hardware SIMD pipeline architecture.
• Metrics including throughput, register pressure, buffer size, and code size are measured and compared across all techniques.
• Experiments are conducted on real benchmarks to quantify trade-offs and validate performance improvements.

Experimental results

Research questions

• RQ1 How does modulo scheduling with unrolling compare to ad hoc and basic modulo scheduling in terms of throughput on SIMD pipelines?
• RQ2 What are the trade-offs between code size, register pressure, and buffer requirements across the three scheduling techniques?
• RQ3 Which scheduling technique achieves the best balance between performance and resource utilization for multimedia streaming workloads?
• RQ4 How does loop unrolling affect the effectiveness of modulo scheduling in reducing initiation interval and improving throughput?
• RQ5 Can the proposed techniques be effectively applied to real-world multimedia benchmarks running on SIMD architectures?

Key findings

• Modulo scheduling with unrolling achieves the highest throughput, reducing the initiation interval by up to 40% compared to basic modulo scheduling in tested benchmarks.
• Ad hoc overlapped execution shows moderate performance gains but incurs higher code size and is less scalable across diverse workloads.
• Register pressure increases significantly with unrolling, but remains within acceptable limits for most multimedia applications.
• Buffer size requirements are minimized with modulo scheduling, while ad hoc scheduling often requires larger buffers due to less optimal data flow management.
• Code size grows with unrolling, but remains manageable and is offset by the performance gains in throughput.
• The study confirms that modulo scheduling with unrolling is the most effective technique for high-throughput multimedia streaming on SIMD pipelines.