ICON: Incast Congestion Control using Packet Pacing in Datacenter Networks
ICON: Incast Congestion Control using Packet Pacing in Datacenter Networks
复制标题
DOI：
10.1109/comsnets.2019.8711175
复制
发表时间：
2019-01
期刊：
影响因子：
--
通讯作者：
Hamed Rezaei;M. Chaudhry;Hamidreza Almasi;Balajee Vamanan
中科院分区：
文献类型：
--
作者：
Hamed Rezaei;M. Chaudhry;Hamidreza Almasi;Balajee Vamanan
数据中心承载着各种应用程序，这些应用程序会产生不同的流量模式，并强加不同的网络目标。在线的、面向用户的应用程序产生多对一的、多为短流的临时流量，这些流量对流完成时间（FCT）的尾部很敏感。数据分析应用程序生成所有对所有流量（例如，网络搜索）的大多数短流饱和网络二分法和作业完成要求所有流完成。后台应用程序（例如，MapReduce）生成大量流，并且由于它们通过网络传输的数据量庞大，因此对吞吐量敏感。虽然数据中心交换矩阵提供良好的二分带宽来处理所有对所有流量，但Incast流量在边缘交换机处受到瓶颈影响，并导致连接到接收器的端口处的队列堆积。由于数据中心交换机使用浅缓冲区来降低成本和延迟，因此由于浅缓冲区容易溢出，从而导致数据包丢失和昂贵的TCP超时，因此队列建立问题进一步加剧。为了解决这些问题，我们提出了ICON，这是一种新颖的方案，它通过调整流量来设置对发送速率的细粒度控制，从而减少了内播引起的数据包丢失。我们提出两个备选方案：简单地在往返时间（RTT）上平滑地对分组进行定步的应用不可知版本和基于应用知识对分组进行定步的应用感知版本（例如，铸造程度）。与现有的最先进的拥塞控制方案相比，ICON实现了短流的第99百分位流完成时间降低77%，长流的吞吐量平均提高18%。此外，ICON将第99百分位数的数据包丢弃平均大幅减少约3倍。
Datacenters host a mix of applications which generate qualitatively distinct traffic patterns and impose varying network objectives. Online, user-facing applications generate many-to-one, incast traffic of mostly short flows, which are sensitive to tail of Flow Completion Times (FCT). Data analytics applications generate all-to-all traffic (e.g., Web search) of mostly short flows that saturate network bisection and the job completions require all flows to complete. Background applications (e.g., MapReduce) generate large flows and are throughput sensitive due to the sheer amount of data that they transfer over the network. While datacenter fabric provides good bisection bandwidth to handle all-to-all traffic, incast traffic is bottle-necked at edge switches and causes queue buildup at the port connected to the receiver. Because datacenter switches use shallow buffers to reduce cost and latency, the queue buildup problem is further exacerbated as the shallow buffers easily overflow causing packet drops and expensive TCP timeouts. To address these issues we propose ICON, a novel scheme which reduces incast-induced packet loss by setting a fine-grained control over sending rate by pacing traffic. We propose two variants: an application agnostic version that simply paces packet smoothly over round trip time (RTT) and an application aware version that paces packets based on application knowledge (e.g., incast degree). Compared to existing state-of-the-art congestion control schemes, ICON achieves 77% lower 99th percentile flow completion times for short flows and 18% higher throughput for long flows on average. Further, ICON drastically reduces 99th percentile packet drops by a factor of about 3 on average.