Estimating Achievable Throughput

A growing number of applications have begun to leverage the flexibility, control, and robustness offered by overlay and grid networks. For example, to achieve better throughput or reliability between a source and destination than the throughput or reliability offered by the standard IP shortest path route, applications have been turning to applicationselected overlay paths. Unfortunately, the underlying Internet over which overlays are constructed offers very little information to help applications make such routing decisions. In particular, it is difficult to obtain accurate information about the path capacity (of “virtual” overlay links), the current traffic level across a path, the available bandwidth along a path, or the expected throughput a transport protocol can expect to get if it were to use the path. Because many applications can benefit from information about the available bandwidth along an overlay path, a variety of available bandwidth estimation techniques have been developed – where available bandwidth is defined as the unused capacity along a path. Examples include Spruce, IGI, Delphi, ABwE, Pathload. A common technique used by many of these tools is the packet pair dispersion technique, in which a sender sends successive probing packets to a receiver using a known time gap between packets and then uses the observed change in the time gap at the receiver to estimate the available bandwidth. The accuracy of this technique depends on whether the queue at the bottleneck router becomes empty between the departure of the first packet and the arrival of the second packet. Thus, several algorithms have been designed that attempted to reduce the impact of invalid probing measurements. Moreover, several available bandwidth estimation techniques rely on knowledge of the path capacity – which must itself be estimated. Although there exists many tools to estimate path capacity, our analysis of these tools shows that