Computational Wireless Network Backplane: Performance in a Distributed Speaker Identification Application Postprint

摘要

A major challenge in the DoD's next-generation network-centric information systems concerns on-demand provisioning of computation and network infrastructures at tactical network edges (e.g., deploying wireless airborne or hybrid air/ground networks). To support this vision, we present DWARF, a general distributed application execution framework for wireless ad-hoc networks which dynamically allocates computation resources and manages failures. DWARF nodes each run a separate task simultaneously, thereby achieving execution speed-up from parallel processing. Failed tasks, e.g., due to fluctuating wireless links to mobile nodes, are automatically detected and reassigned, transparent to the application. Further, tasks are executed in an order that satisfies dependencies given by task dependency graphs. To use DWARF, application programmers need only decompose their applications into tasks and define the task dependency graphs. In this paper, we describe DWARF and report its benefits in running an important existing application, speaker identification over a 32-node wireless network which supports fault-tolerant computation. We observed two major performance gains: (1) a ten-fold speed-up in identifying speakers due to parallelizing the application, and (2) higher accuracy in speaker identification, made possible by the increased sensor diversity provided by geographically distributed nodes. While our nodes have modest computing power individually, combined under DWARF, they are able to execute speaker identification with much greater speed and with improved accuracy.