Incorporating change detection in network coordinate systems for large data transfers

The performance of large scale applications, such as those enabled by service-oriented, grid and cloud technologies, heavily relies on aspects related to the network topology and latency. As such, predicting the actual communication latencies is of high interest. The current state-of-the-art solutions to the problem of estimating the latency among distributed nodes comprise algorithms that tend to build upon the notion of network coordinates (NCs). Since network conditions change continuously, NCs need to be updated very frequently and, thus, are prone to oscillations. We present a variant of the pioneer NC algorithm, called Vivaldi, which encapsulates a change detection mechanism to prohibit NCs updates unless the network conditions change significantly. The contribution of this paper is twofold: first, to assess the impact of change detection and, second, to evaluate the NC algorithms in a realistic service-based environment where real measurements refer to large data transfers, contrary to current approaches that collect feedback from much smaller data transmissions, such as pings. The evaluation shows that our variant improves both performance and stability with less overhead.