The Longcut Wide Area Network Emulator. Design and Evaluation

Abstract

Experiments run on a Grid, consisting of clusters administered by multiple organizations connected by shared wide area networks (WANs), may not be reproducible. First, traffic on the WAN cannot be controlled. Second, allocating the same resources for subsequent experiments can be difficult. Longcut solves both problems by splitting a single cluster into several parts, and for each part having one node emulating a WAN link by delaying messages sent through it. The delay is calculated using latency and bandwidth measurements collected using the Network Weather Service and a parallel application monitor. We evaluate the precision, usability for WAN collective operation research, and scalability of Longcut.