A Fast and Accurate Technique for Mapping Parallel Applications on Stream-Oriented MPSoC Platforms with Communication Awareness

Abstract

The problem of allocating and scheduling precedence-constrained tasks on the processors of a distributed real-time system is NP-hard. As such, it has been traditionally tackled by means of heuristics, which provide only approximate or near-optimal solutions. This paper proposes a complete allocation and scheduling framework, and deploys an MPSoC virtual platform to validate the accuracy of modelling assumptions. The optimizer implements an efficient and exact approach to the mapping problem based on a decomposition strategy. The allocation subproblem is solved through Integer Programming (IP) while the scheduling one through Constraint Programming (CP). The two solvers interact by means of an iterative procedure which has been proven to converge to the optimal solution. Experimental results show significant speed-ups w.r.t. pure IP and CP exact solution strategies as well as high accuracy with respect to cycle-accurate functional simulation. Two case studies further demonstrate the practical viability of our framework for real-life applications

A modular fast simulation framework for stream-oriented MPSoC

“…"penalty value" and "exponential function". We argue that neither case properly reflects actual behavior, because the throughput is not always functional to the bus bandwidth, as shown in [16]. In our framework the communication is a first class citizen with assigned arbiters which properly schedule the traffic and therefore can reflect contentions.…”

Robust Scheduling of Task Graphs under Execution Time Uncertainty

“…Even though the allocation and scheduling problems remains NP-Hard, efficient complete (exact) algorithms are known (see for instance [11] and references therein) that work well in practice; many incomplete (heuristic) approaches have been proposed for very large problems that exceed the capabilities of complete solvers [12]. These approaches can also be used in the case of variable execution times, but they need to force determinism: at run-time, task execution can be stretched artificially (e.g.…”

“…In details, two set of instances (referred to as Group 1 and Group 2 in the following) were obtained by means of a flexible generator, designed to produce realistic Task Graphs 11 . Group 1 consists of TGs containing 20 to 70 nodes prior the mapping process; the number of outgoing arcs for each node (branching factor) is in the range 3-5 (sink nodes excluded); TGs in group 2 have fixed number of nodes prior to mapping (namely, 40) and branching factor from 2-4 to 6-8.…”

“…Obviously, if lb = ub 10 The production and characterization of a sufficient number of real applications would take prohibitive time. 11 The generator has been presented at the PST ICAPS2007 workshop and is available for download at http://ai.unibo.it/node/410. the tightest deadline has been identified.…”

Reducing the Abstraction and Optimality Gaps in the Allocation and Scheduling for Variable Voltage/Frequency MPSoC Platforms

“…Finally, we intended to compare the computation efficiency of our hybrid approach with that of traditional approaches not leveraging problem decomposition (i.e., the whole mapping problem modeled through IP or CP). However, such a comparison was already reported in [9] for a simpler problem (power consumption was not accounted for, and only pipelined task graphs were supported), where a computation efficiency gap of orders of magnitude was already shown. Considering an upper bound of 15 min for the search time, CP and IP proved capable of finding the optimal solution only for extremely small instances, with a low number of tasks and processing elements, and of finding a solution (not the optimal one) only in 50% of the hard instances, while the hybrid approach solved 100% of the instances to optimality.…”