DMCA
NEPAL: A Framework for Efficiently Structuring Applications for Network Processors (2003)
| Citations: | 14 - 0 self |
BibTeX
@MISC{Memik03nepal:a,
author = {Gokhan Memik and William H. Mangione-Smith},
title = {NEPAL: A Framework for Efficiently Structuring Applications for Network Processors},
year = {2003}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
Network Processors are becoming increasingly common for implementing processing elements in networks. In this worl we propose a new design environment that increases system performance by revealing program structure. The design environment assumes that problems such as task communication and synchronization will be handled by hardware mechanisms. Therefore, the tasks for the developer are significantly reduced. This abstraction is achieved by generating an execution tree that is processed at run-time by a dynamic manager. The tree consists of modules, which are highlevel work units for many networking applications. Considering the module tree, the dynamic module manager decides which modules to activate and selects an appropriate execution core. We use two heuristics to guide dynamic task allocation. The first tries to achieve a load with minimal variance, and the second tries to minimize the inter-core communication. We show that network processors with different structures can take advantage of this modularity. Particularly, we show that the execution time of a set of representative applications is reduced by as much as 62.4% (36.2% on average) for a processor utilizing a global shared bus for communication, and by as much as 61.1% (34.4% on average)for a processor that has special neighbor-to-neighbor communication. We further show that the modular representation of the applications might also be used for task level speculative execution, increasing the performance of a single-core processor by 43.1% on average.
Keyphrases
network processor efficiently structuring application first try different structure task communication highlevel work unit hardware mechanism new design environment program structure minimal variance single-core processor module tree system performance task level speculative execution dynamic task allocation second try design environment many networking application modular representation execution tree tree consists dynamic manager execution time dynamic module manager inter-core communication processing element special neighbor-to-neighbor communication appropriate execution core representative application
