Comparison of Execution Paradigms

The decoupled execution paradigm reshapes the conventional HEC execution paradigm of “retrieve - compute - store” cycles into “retrieve - reduce - compute - reduce - store” cycles as shown in the Figure below, where the “reduce” phases are designed to conduct data-intensive operations and reduce data size before moving data across the network. These retrieval, reduce, compute, and store phases can be pipelined to overlap the I/O, communication, and computation times. From one point of view, DEP is an enhanced version of MapReduce, where the “reduce” is not conducted by one node with its local storage, but a set of (data) nodes and the global storage, so that parallel computing features can be maintained. From another point of view, the data nodes are the data-access accelerators, to speed up the storage data-access delay and reduce data size before sending data across the network.

dep-compare-web.jpg

Recent changes RSS feed Debian Powered by PHP Valid XHTML 1.0 Valid CSS Driven by DokuWiki