Seminar by Dr. Mukul Goyal
Predicting TCP Throughput from Non-invasive Network Sampling
Dr. Mukul Goyal
Department of Computer and Information Science
Ohio State University
Columbus, Ohio, USA
Date: Monday, August 4, 2003
Time: 5:00 PM
Venue: CS-101
Abstract
As the Internet continues to evolve into the dominant commercial communications infrastructure, the need for service verification and quality monitoring is also increasing. The conversion of network level performance characteristics into representative user and application level performance metrics is an open research problem. As TCP traffic represents the bulk of data traffic on the Internet, the average throughput achieved by a long life TCP flow is an important user level performance metric. In our research, we have focussed on understanding how well can we translate the available network performance data into the average throughput achieved by a long life TCP flow. This problem can essentially be divided into two sub-problems: 1) How well can we estimate network performance charactersitics like loss rate and delay observed by individual TCP flows in a "non-invasive" and "scalable" manner? 2) How well can we predict long life TCP throughput based on non-invasive estimates of network performance characteristics? Via extensive simulations and testbed experiments, we examine the suitability of available non-invasive and scalable methods of estimating network performance characteristics. This is followed by the derivation and thorough performance analysis of a new analytical model that uses non-invasively determinable network performance characteristics to predict the average throughput of long life TCP flows.