In the context of large-scale datacenter networks, the distinct mechanisms of these two protocols result in mutual interference when TCP and RoCE flows traverse shared segments of network. TCP may preempt the bandwidth allocated to RoCE, leading to unfair allocation of shared buffers. Nevertheless, due to the cost trade-off, the consideration of shared buffers becomes necessary. Therefore, isolating through traffic classes does not effectively address the issue. To meet the challenge, in this paper, we present BufferConcede, a switch-based traffic control method that allows TCP traffic to occupy more buffer space in non-hybrid switch, thereby weakening TCP traffic’s greediness for buffer space. We integrate it with practical congestion control approaches such as DCQCN, DCTCP. We evaluate BufferConcede in both implementation and large-scale simulations. Compared with state of the act, BufferConcede can reduce bandwidth allocation errors under 2.0% and reduce the average FCT of RoCE traffic by approximately 4.8%–10.2%.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

BufferConcede: Conceding Buffer for RoCE Traffic in TCP/RoCE Mix-Flows

  • Lingxuan Meng,
  • Kaiyun Liu,
  • Weibei Fan,
  • Fu Xiao,
  • Mengjie Lv,
  • Lei Han,
  • Xiaoyan Li

摘要

In the context of large-scale datacenter networks, the distinct mechanisms of these two protocols result in mutual interference when TCP and RoCE flows traverse shared segments of network. TCP may preempt the bandwidth allocated to RoCE, leading to unfair allocation of shared buffers. Nevertheless, due to the cost trade-off, the consideration of shared buffers becomes necessary. Therefore, isolating through traffic classes does not effectively address the issue. To meet the challenge, in this paper, we present BufferConcede, a switch-based traffic control method that allows TCP traffic to occupy more buffer space in non-hybrid switch, thereby weakening TCP traffic’s greediness for buffer space. We integrate it with practical congestion control approaches such as DCQCN, DCTCP. We evaluate BufferConcede in both implementation and large-scale simulations. Compared with state of the act, BufferConcede can reduce bandwidth allocation errors under 2.0% and reduce the average FCT of RoCE traffic by approximately 4.8%–10.2%.