A Case for Queue-to-Queue, Back-Pressure-Based Congestion Control for Grid Networks.
Marc Herbert, Pascale Vicat-Blanc Primet · Parallel and Distributed Processing Techniques and Applications · 2004
Standard “new-Reno” TCP faces some performance limitations in very high throughput IP WAN networks, (e.g., computing grids) due to a long end-to-end congestion feedback loop and a conservative behaviour with respect to congestion. We investigate breathing new life into the following idea: implement aggressive congestion control via hop-by-hop, link-level flow control (back-pressure). Studying TCP’s behaviour at the packet level on an emulated WAN link allows us to explain: a) the dependence between TCP performance and burstiness at the Ethernet level; b) the performance and fairness gains obtained by the activation of Ethernet 802.3x flow control. Recognising that data moves not from hop to hop but from queue-to-queue, regardless of the layer, we propose to refine previous hop-by-hop approaches to a new “Network of Queues” (NoQ) approach, where each network buffer protects itself against overflow, thus preventing global network congestion. We finally identify advantages and limitations and estimate implementation costs of this approach.