Meta-Algorithms for Scheduling a Chain of Coarse-Grained Tasks on an Array of Reconfigurable FPGAs

Abstract

This paper considers the problem of scheduling a chain of n coarse-grained tasks on a linear array of k reconfigurable FPGAs with the objective of primarily minimizing reconfiguration time. A high-level meta-algorithm along with two detailed meta-algorithms (GPRM and SPRM) that support a wide range of problem formulations and cost functions is presented. GPRM, the more general of the two schemes, reduces the problem to computing a shortest path in a DAG; SPRM, the less general scheme, employs dynamic programming. Both meta algorithms are linear in n and compute optimal solutions. GPRM can be exponential in k but is nevertheless practical because k is typically a small constant. The deterministic quality of this meta algorithm and the guarantee of optimal solutions for all of the formulations discussed make this approach a powerful alternative to other metatechniques such as simulated annealing and genetic algorithms.

Combining Checkpointing and Replication for Reliable Execution of Linear Workflows

“…In other words, the task graph is a linear chain, and each task (except maybe the first one and the last one) reads data from its predecessor and produces data for its successor. Such linear chains of tasks also appear in image processing applications [25], and are usually called linear workflows [34].…”

“…06 4 0. 25 (b) αi = 0.5 Figure 9: Impact of checkpoint/recovery cost and error rate on the usage of checkpointing and replication. Total work is fixed to 10, 000s, there are n = 20 tasks and the work is distributed according to the Decreasing distribution.…”