Towards run-time actor mapping of dynamic dataflow programs onto multi-core platforms

Abstract

The emergence of massively parallel architectures, along with the necessity of new parallel programming models, has revived the interest on dataflow programming due to its ability to express concurrency. Although dynamic dataflow programming can be considered as a flexible approach for the development of scalable applications, there are still some open problems in concern of their execution. In this paper, we propose a low-cost mapping methodology to map dynamic dataflow programs over any multi-core platform. Our approach finds interesting mapping solutions in few milliseconds that makes it doable at regular time by translating it in an equivalent graph partitioning problem. Consequently, a good load balancing over the targeted platform can be maintained even with such unpredictable applications. We conduct experiments across three MPEG video decoders, including one based on the new High Efficiency Video Coding standard. Those dataflow-based video decoders are executed on two different platform: A desktop multi-core processor, and an embedded platform composed of interconnected and tiny Very Long Instruction Word -style processors. Our entire design flow is based on open-source tools. We present the influence of the number of processors on the performance and show that our method obtains a maximum decoding rate for 16 processors.

Embedded Multi-Core Systems Dedicated to Dynamic Dataflow Programs

“…Therefore, the maximum speedup achieved with our embedded implementation is much bigger than the maximum speedup achieved with the implementation on general-purpose processors (which seems to be around 3x [31]). This can be mainly explained by the fact that the communications between the cores within our embedded implementation do not induce any overhead compared to more conventional communication and memory schemes implemented in generalpurpose processors.…”

“…In previous works [32,31], we have introduced run-time actor mapping/scheduling strategies dedicated to DPN-based actors. Our scheduling strategies execute the current actor until it cannot fire anymore to exploit spatial and temporal locality.…”

“…Indeed, the number of processors available must be limited so as to reduce the power consumption of the platform. As opposed to the previous experimentation, the actors are mapped by an automated system [31] which takes into account the irregularity of our applications thanks to a profiling step, as presented in Figure 8. Our mapping system starts by analyzing the communication rates and the computational loads, as we did respectively in Section 5.2 and Section 5.3.…”

DKPN: A Composite Dataflow/Kahn Process Networks Execution Model

“…They also present a distributed scheduler with static mapping. In subsequent work [20], they enable automated mapping of actors through profiling, but still not dynamically at run time. In their latest paper [21], they target specifically MPSoCs but the scheduling and mapping techniques remain the same.…”

Efficient software synthesis of dynamic dataflow programs

“…In previous works [13,14], we have introduced run-time actor mapping/scheduling strategies dedicated to DPN-based actors. Our scheduling strategies execute the current actor until it cannot fire anymore to exploit spatial and temporal locality.…”