Planned Direct Transfers: A Programming Model for Real-Time Applications Mercury Backgrounder - June 15, 1996

This paper presents the "Planned Direct Transfer" programming model, developed by Mercury Computer Systems to meet the requirements of embedded high-performance computing applications. In this model, data transfers are "Planned" before they occur, resulting in low software overhead execution; they are also "Direct"&emdash;they do not require intermediate data copying. This paper locates the Planned Direct Transfer (PDT) model in the landscape of the standard approaches of Shared Memory and Message Passing. 1: PDT: Key Objectives and Assumptions The Planned Direct Transfer (PDT) programming model strives to contain the software component of interprocessor communication management within the initialization phase of parallel programs, thus letting the hardware take care of the data movements during the main phase of the program execution. That way, all the computational and resource allocation overhead associated with communications can be incurred before the transfers actually take place. The PDT model is not universal; it relies on two assumptions. First, PDT was developed and exercised on a specific class of applications, those requiring high-performance large-scale real-time computing. These real-time applications (RTAs) typically involve embedded sensor-based signal or image processing, and are exemplified by Synthetic Aperture Radar (SAR), Space-Time Adaptive Processing (STAP), Adaptive Beamforming, and security and medical imaging. Secondly, like all programming models, PDT makes specific assumptions on the architecture and hardware it runs on. In particular, PDT assumes high-bandwidth interconnect fabric as well as low-latency direct read and write capability between distributed memory modules, with no intermediate data copying. The PDT model was developed by Mercury Computer Systems on the RACE® architecture [1,2]. 2: Specific Demands of Real-time Applications RTAs are characterized by demands that are distinct from those of traditional scientific applications. Some of these demands are more exacting, some are more lenient. The more exacting demands require highly tuned resources and handles for optional programmer low-level fine tuning. The relative leniency of some RTA's demands explain how PDT can exceed the performance of most supercomputers at a fraction of their cost. 1 of 6 5/1/01 12:56 PM Planned Direct Transfers http://www.mc.com/backgrounder_folder/IPPS-PDT/ipps_pdt.htm