It is now possible to provide on-demand guaranteed bandwidth between end systems by leasing wavelengths from lambda-grid networks. With the availability of 10 Gig Network Interface Cards (NICs), an end-to-end dedicated connection of 10 Gbps may be established. Although physical connectivity is available, the bottleneck for data transfer often turns out to be the end system performance. This bottleneck may be at any one of the following steps: (i) data transfer from the NIC to the socket, (ii) the application reading data from the socket, or (iii) data transfer to other sub-systems such as the disk, graphics card, etc. In the absence of definitive knowledge about the bottleneck rate in the receiving end system, the sender’s transmission rate often oscillates between extremes because it overshoots the critical bottleneck rate of the receiver; this typically generates a significant number of negative acknowledgments and hence quenches the data rate, resulting in poor performance such as larger file transfer times, and poor jitter.
We believe that in order to optimize the performance of the transport protocols and achieve the important flow control functionality, it is important to estimate the end system effective bottleneck rate. We propose a method of modeling and active analysis of the end system to estimate the effective bottleneck rate. We propose to use a queueing network model that incorporates the underlying system resources as well as the competing processes executing at the end system. The input parameters of the model such as the processing rates may be determined by a monitoring tool and the solution of the model provides the bottleneck rate at the end system. The goal of the proposed research is to demonstrate the effectiveness of this approach and subsequently develop a software which will interface with existing transport protocols and provide the receiving end system bottleneck rate to the sending end system. The sender will match the sending rate to the bottleneck rate and thereby improve the performance of data transfer.
The technical merits of the proposed research are the following:
1. We will develop queueing network models for representing the end system in different scenarios. We will investigate tools which may be used to timestamp kernel level events and determine the input parameters to the queuing network model. We will develop an approach to determine the effective bottleneck rate given the current workload of the end system from the above queuing model. 2. We will develop software that easily integrates with existing transport protocols and delivers the effective bottleneck rate determined by the to the sender. This software must be easily deployable. It must require minimum manual customization. 3. We will carry out experimental analysis for different end system applications and for different end system configurations. Through our experimental studies, we will evaluate and refine the proposed software. 4. We will demonstrate that our approach improves the performance of end-to-end transfer in a variety of settings.
The broader impacts of the proposed work are the following: 1. Our software will be made available to the research community to analyze and optimize similar applications and systems. 2. The methods developed will be applicable to other systems such retrieving data from network storage and downloading of video for Video-on-Demand (VoD) services in an IP television (IPTV) network. 3. The research project will provide a framework to train graduate and undergraduate students in both analytical and experimental methods and develop knowledge and intuition in next generation computer systems and distributed applications.
Dipak Ghosal Matt Farrens Biswanath Mukherjee Vishal Ahuja Rennie Arvhibals Amitabha Banerjee Andrei Dragos