End-to-End Dataflow Parallelism for Transfer Throughput Optimization

Introduction

The data transfer throughput is a major factor that affects the performance of applications from many scientific areas (e.g. high-energy physics, bioinformatics, numerical relativity and computational fluid dynamics). The advancements in optical networking technology have gone beyond the achievable throughput values the applications get, however the same speed up is not seen in the application performance due to many reasons such as the protocol inadequacy, poorly tuned protocol parameters and underutilized capacity of the end-systems. The current protocols that are highly common (e.g. TCP) were originally not designed for high-bandwidth networks. Due to its additive increase multiplicative decrease policy, TCP takes a long time to fill the pipe of long-fat network pipes. Many protocols have been designed for high-bandwidth networks in the transport layer (Kola & Vernon, 2007; Jin et al, 2005; Floyd, 2003) to overcome this problem however they fail to replace TCP.

Other than transport layer protocols, some application-level solutions are proposed as well. Two of the very common methods are tuning buffer size and using parallel streams. While some buffer tuning methods need modification to the kernel (Cohen &Cohen, 2002; Semke, Madavi & Mathis, 1998; Torvalds et al, 2010; Weigle & Feng, 2001), the others are done at the application level (Jain, Prasad & Davrolis,2003; Prasad, Jain &Davrolis, 2004, Hasegawa et al 2001; Morajko, 2004). Although the buffer size parameter is properly tuned, it does not show a better performance than using parallel streams because parallel streams recover from packet loss quickly rather than a single stream with tuned buffer. They achieve high throughput by mimicking the behavior of individual streams and get an unfair share of the available bandwidth (Sivakumar, Bailey & Grossman, 2000; Lee et al, 2001; Balakrishman et al, 1998; Hacker, Noble & Atley, 2005; Eggert, Heideman & Touch, 2000; Karrer, Park & Kim, 2006; Lu, Qiao & Dinda, 2005). However excessive usage of parallel streams reaches the network to a congestion point and it is hard to predict this point. The studies that try to find the optimal number of streams are so few and mostly based on approximate theoretical models (Hacker, Noble & Atley, 2002; Lu et al, 2005; Altman et al, 2006; Kola & Vernon, 2007}. They all have specific constraints and assumptions. Also the correctness of the proposed models is mostly proved with simulation results.

The foretold solutions to improve the throughput only remove the disadvantages of the protocols used. However, at some point the end-system resources become the source of bottleneck such as CPU, disk and NIC itself. Additional parallelism is needed through striping but the optimal level of striping is an open research area. The existing tools such as the GridFTP striped server (Allcock et al, 2005) and Dmover (Nathan et al, 2010) provide a means to utilize striping through multiple CPUs and nodes of an end-system architecture but they give the preference to the user to define the parallelism level. A dynamic and autonomic system that will decide this level depends on many factors.

In this book chapter we discuss many factors that affect the end-to-end application throughput such as the buffer size, parallel streams, CPU speed, disk speed and access methods in systems that use high-speed networks. The major purpose of this chapter is to provide insight to the characteristics of the end-systems that cause the bottleneck for throughput and to discuss future directions. We also present a method to optimize the parallel stream number and we have seen that this model gives very accurate results regardless of the type of the network.