Performance analysis of different TCP variants in wireless ad hoc networks

End-to-End TCP Congestion Control for Mobile Applications

“…It is increasingly acknowledged that TCP relying on packet loss as congestion measure is not stable and efficient, especially for wireless networks. This motivates a number of TCP schemes that employ queueing delays as congestion measure to implement congestion control [8], [11], [12]. In order to achieve the desired optimization, congestion control algorithms in Mo-walrand scheme and Fast-TCP [4], [29] were proposed by using queueing delays as Lagrange multipliers.…”

“…where (11) is due to the fact that the flow rate x s is equal to the window size w s divided by the total RTT d s + q s of flow s; (13) and (15) are due to link capacity constraints; (12) and (14) are due to the the following reasons: because the data packet can be divided indefinitely (i.e., infinitely small), when the total rate through the link l is less than its capacity c l or r l , the queue size (and queueing delay) on this link is zero.…”

“…Properties 2) and 3) are new. Property 2) can be derived by examining the partial first derivative ∂x s /∂w s , whereas property 3) is a consequence of property 2) and equation (11).…”

“…The marginal increase ∂x s /∂w s is at most 1/d s which occurs when its aggregate queueing delay q s = 0 [cf. (11)], i.e., the network is under-loaded. Note that although the aggregate queueing delay vector q s is uniquely determined for a given w, the queueing delay vector q is not necessarily unique, unless the number of independent bottleneck links is equal to the number of sources such that q can be solved from q s .…”

“…The BS implements the scheduling policy (10). Taking into account the feedback delay, we run TCP-Reno [11], TCP-Vegas [8], TCP-Cubic [20], TCP-Compound [19], FAST-TCP [4], and the proposed QUIC-TCP (ρ = 0.5). The window update step size for FAST-TCP and QUIC-TCP is set as κ = 1.5.…”