Aggregating the bandwidth of multiple low-cost physical communication links to form a single, higher capacity logical link is a common and inexpensive approach to increasing network performance. Aggregating or bonding links has traditionally been used in relatively static settings such as LAN-WAN interconnection, where the component links to be aggregated were typically identical (e.g., multiple T1 links) with very similar communication characteristics. However the emergence of multi-homed hosts and diverse access technologies (e.g., cable modem, wireless DSL) compel us to study the performance of aggregating the bandwidth of increasingly heterogeneous links.In this paper we examine the end-to-end performance of a single TCP connection with packets transmitted over multiple heterogeneous links. We describe a testbed we constructed, present empirical results, and introduce a new Linux-based traffic scheduler called wTEQL that we have developed to optimize the performance of TCP over multipleheterogeneous communication links. Finally, we present a novel analytical approach based on linear systems theory to characterize the performance of TCP over an inverse multiplexed channel with WRR scheduling over heterogeneous links. We believe that our model is the first to demonstrate analytically how TCP throughput on two links with a poorly matched packet scheduling policy can fall below the achievable throughput of the slower of the two links when that link is used alone.1展开▼
