Lossy Bulk Synchronous Parallel Processing Model for Very Large Scale Grids

TL;DR

This paper explores a lossy bulk synchronous parallel (BSP) model for large-scale grid computing, demonstrating that strategic packet duplication over UDP can mitigate packet loss effects and improve parallel performance.

Contribution

It introduces a BSP-based model that incorporates packet duplication strategies to enhance performance in high-loss network environments.

Findings

Packet loss rates of 5%-15% are common between Internet-distributed nodes.

Transmitting multiple copies of packets increases speedup in parallel processing.

Optimal packet duplication depends on the number of grid nodes to maximize performance.

Abstract

The performance of a parallel algorithm in a very large scale grid is significantly influenced by the underlying Internet protocols and inter-connectivity. Many grid programming platforms use TCP due to its reliability, usually with some optimizations to reduce its costs. However, TCP does not perform well in a high bandwidth and high delay network environment. On the other hand, UDP is the fastest protocol available because it omits connection setup process, acknowledgments and retransmissions sacrificing reliable transfer. Many new bulk data transfer schemes using UDP for data transmission such as RBUDP, Tsunami, and SABUL have been introduced and shown to have better performance compared to TCP. In this paper, we consider the use of UDP and examine the relationship between packet loss and speedup with respect to the number of grid nodes. Our measurement suggests that packet loss rates between 5%-15% on average are not uncommon between PlanetLab nodes that are widely distributed over the Internet. We show that transmitting multiple copies of same packet produces higher speedup. We show the minimum number of packet duplication required to maximize the possible speedup for a given number of nodes using a BSP based model. Our work demonstrates that by using an appropriate number of packet copies, we can increase performance of parallel program.