AMP: an adaptive multipath TCP for data center networks

Abstract

In recent years several multipath data transport mechanisms, such as MPTCP and XMP, have been introduced to effectively exploit the path diversity of data center networks (DCNs). However, these multipath schemes have not been widely deployed in DCNs. We argue that two key factors among others impeded their adoption: TCP incast and minimum window syndrome. First, these mechanisms are ill-suited for workloads with a many-to-one communication pattern, commonly found in DCNs, causing frequent TCP incast collapses. Second, the syndrome we discover for the first time, results in 2-5 times lower throughput for single-path flows than multipath flows, thus severely violating network fairness.To effectively tackle these problems, we propose AMP: an adaptive multipath congestion control mechanism that quickly detects the onset of these problems and transforms its multipath flow into a single-path flow. Once these problems disappear, AMP safely reverses this transformation and continues its data transmission via multiple paths. Our evaluation results under a diverse set of scenarios in a fat-tree topology with realistic workloads demonstrate that AMP is robust to the TCP incast problem and improves network fairness between multipath and single-path flows significantly with little performance loss.

AMP: An Adaptive Multipath TCP for Data Center Networks

“…We also test those schemes with realistic web search and data mining workloads [23] and under a more intensive and dynamic condition with an order of magnitude shorter flow arrival rate in a 4:1 oversubscribed fat-tree topology. We make similar observations across all cases (an extended version of this paper is available in [9] that includes all these results).…”

“…We thus set τ = 8 as default (to be conservative). The exact graph has been omitted for brevity, but it is available in [9].…”

“…We examine the performance of existing works such as eXplicit MultiPath (XMP) [4] and Data Center MultiPath (DCM) [9] (our prior work), which follow such an integration. It turns out that these provide fast load balancing while keeping switch buffer occupancy low.…”