A flow‐executing scheme for DOACROSS loops on dynamic dataflow machines

Abstract

AbstractThis paper modifies a flow‐executing scheme of the color‐reuse type, using multiple initial loop control packets, and then proves that the flow‐executing scheme is best suited for executing DOACROSS loops on dynamic dataflow machines. Flow‐executing schemes can be divided into four categories: (1) those using a single initial loop control packet; (2) those using multiple initial loop packets; (a) the color overflow type; and (b) the color reuse type. Then the flow‐executing scheme can be classified into Classes (1‐a), (1b), (2‐a), and (2‐b) through the combination of Categories (1), (2), (a), and (b). This paper suggests that Class (2‐b) is best suited for executing DOACROSS loops, as it extracts full parallelism from DOACROSS loops, no sychronization overhead exists, and no memory access overhead exists after the synchronization.