Oscillations of the Sending Window in Compound TCP

Abstract

Abstract. One of the key ideas of Compound TCP is to quickly increase the sending window, until full link utilization is detected, and then to keep it constant for a certain period of time. The actual Compound TCP algorithm does not hold the window constant but, instead, it makes it oscillate around the desired value. Using an analytical model and ns-2 simulations we study these oscillations on a Linux implementation of Compound TCP, in the case of a single connection with no cross tra c. Even in this simple case we show how these oscillations can behave in di erent ways depending on the bandwidth delay product. We also show how it is important to take into account, in the analytical model, that some implementation subtleties may introduce non-negligible di erences in the behavior of the protocol.

Compound TCP with Random Losses

“…In such case the window in phase 2 is equal to θ μτ +γ, where μ is the capacity of the bottleneck link andτ is the round trip propagation delay. In reality any queue size estimate available at the sender is outdated due to feedback delays, this fact combined with the CTCP algorithm presented in [7] causes the window to oscillate during this phase as we analyzed in [17]. The length of phase 2 is dictated by the "congestion component" of the window.…”

Modelling an Isolated Compound TCP Connection

“…To the best of our knowledge there are no complete theoretical models of CTCP that can be used to analyze in details the behavior of this new protocol. While we have shown in [5] that the sending window oscillates during this phase, and that these oscillations may have a significant impact on the performance of CTCP. Other evaluations are based on experiments.…”

“…As noted in [5], during phase 3, the algorithm described in [14] and [12] causes the window to oscillate around a constant value. Depending on the system parameters it is possible to have different patterns.…”

“…While it is possible to use a fluid model to estimate the size of the oscillations we believe it is easier to use the discrete event model presented in [5] to precisely characterize these oscillations. In the remainder of this section, after a brief presentation of the model used in [5], we will extend that work and by showing how it is possible to determine the m:n pattern of the oscillations based on the bandwidth delay product.…”

“…In the remainder of this section, after a brief presentation of the model used in [5], we will extend that work and by showing how it is possible to determine the m:n pattern of the oscillations based on the bandwidth delay product.…”