Gentle Slow Start to Alleviate TCP Incast in Data Center Networks

Abstract

Modern data center networks typically adopt symmetric topologies, such as leaf-spine and fat-tree. When a large number of transmission control protocol (TCP) flows in data center networks send data to the same receiver, the congestion collapse, called TCP Incast, frequently happens because of the huge packet losses and Time-Out. To address the TCP Incast issue, we firstly demonstrate that adjusting the increasing speed of the congestion window during the slow start phase is crucially important. Then we propose the Gentle Slow Start (GSS) algorithm, which adjusts the congestion window according to real-time congestion state in a gentle manner and smoothly switches from slow start to congestion avoidance phase. Furthermore, we present the implementation and design of Gentle Slow Start and also integrate it into the state-of-the-art data center transport protocols. The test results show that GSS effectively decreases the Incast probability and increases the network goodput by average 8x.

Latency and Throughput Optimization in Modern Networks: A Comprehensive Survey

“…Despite the previous studies to find the maximum possible startup rate, there are studies tried to reduce the startup rate. Gentle slow start [277] is a method that tries to alleviate the incast problem in DC networks.…”

Fine-Grained Congestion Control for Multipath TCP in Data Center Networks

“…For saving cost, these switches normally have small-size SRAM packet buffers. When the switch buffers are shared by both large and small flows, it is very common that the small switch buffer is occupied by a few number of large flow, leading to large queuing delay and TCP incast problem [11]- [14]. In order to satisfy the latency requirement of small flows, the low buffer occupation should be achieved.…”