Analysis for TCP in data center networks: Outcast and Incast

摘要

The unfairness caused by bandwidth sharing via TCP in data center networks is called TCP Outcast problem. Some researchers show that the throughput of a flow with small Round Trip Time (RTF) is less than that with large KIT which is completely contrary to the classic Transmission Control Protocol (TCP) protocol, and they believe that the Outcast problem is caused by port blackout in data center. However, we find that there are more important reasons for outcast which are an unfair distribution of the flows with different RTT on physical link and the differences in congestion window size when finishing the transmission of current block. TCP Incast is a throughput collapse that occurs when many flows arrive at the switch destined for the same output port. Incast deteriorates the performance of networks by increasing the queuing delay of flows and decreasing the throughput of applications. In this paper, we develop an analytical model for goodput when TCP Outcast occurs. Then we propose a new protocol based on window notification mechanism called TCP congestion window replacement (TCP-CWR), which can improve the goodput of the flows with small RTT to solve the outcast problem. Second, we propose a new transport layer protocol based on acknowledgment (ACK) reply changing rate to solve the Incast problem named TCP with Acknowledgment Changing Rate (TCP-ACR). The TCP-ACR protocol adjusts the current congestion window according to the changing rate of ACK and the theoretical maximum congestion window, and it can alleviate the congestion in data center networks. Last, we conduct simulation experiments to verify our proposed schemes. The results show that our analysis of Outcast is correct and our proposed protocols for Outcast and Incast are effective and practical. (C) 2016 Elsevier Ltd. All rights reserved.