For recent datacenter networks, RDMA (Remote Direct Memory Access) can ease the overhead of the TCP/IP protocol suite. The RoCEv2 (RDMA over Converged Ethernet version 2) standard enables RDMA on widely deployed Ethernet technology. RoCEv2 leverages priority-based flow control (PFC) for realizing the lossless environment required by RDMA. However, PFC is well-known to have the technical weakness of head-of-line blocking. Congestion control for RDMA is a very hot research topic for datacenter networks. In this paper, we propose a novel congestion control algorithm for RoCEv2, TIDD (Timer-based Increase and Delay-based Decrease). TIDD basically combines the timer-based increase of DCQCN and delay-based decrease of TIMELY. Extensive simulation results show that TIDD satisfies the high throughput and low latency required for datacenter networks.
Masahiro NOGUCHI
Kansai University
Daisuke SUGAHARA
Kansai University
Miki YAMAMOTO
Kansai University
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Masahiro NOGUCHI, Daisuke SUGAHARA, Miki YAMAMOTO, "Timer-Based Increase and Delay-Based Decrease Algorithm for RDMA Congestion Control" in IEICE TRANSACTIONS on Communications,
vol. E105-B, no. 4, pp. 421-431, April 2022, doi: 10.1587/transcom.2021WWP0005.
Abstract: For recent datacenter networks, RDMA (Remote Direct Memory Access) can ease the overhead of the TCP/IP protocol suite. The RoCEv2 (RDMA over Converged Ethernet version 2) standard enables RDMA on widely deployed Ethernet technology. RoCEv2 leverages priority-based flow control (PFC) for realizing the lossless environment required by RDMA. However, PFC is well-known to have the technical weakness of head-of-line blocking. Congestion control for RDMA is a very hot research topic for datacenter networks. In this paper, we propose a novel congestion control algorithm for RoCEv2, TIDD (Timer-based Increase and Delay-based Decrease). TIDD basically combines the timer-based increase of DCQCN and delay-based decrease of TIMELY. Extensive simulation results show that TIDD satisfies the high throughput and low latency required for datacenter networks.
URL: https://global.ieice.org/en_transactions/communications/10.1587/transcom.2021WWP0005/_p
Copy
@ARTICLE{e105-b_4_421,
author={Masahiro NOGUCHI, Daisuke SUGAHARA, Miki YAMAMOTO, },
journal={IEICE TRANSACTIONS on Communications},
title={Timer-Based Increase and Delay-Based Decrease Algorithm for RDMA Congestion Control},
year={2022},
volume={E105-B},
number={4},
pages={421-431},
abstract={For recent datacenter networks, RDMA (Remote Direct Memory Access) can ease the overhead of the TCP/IP protocol suite. The RoCEv2 (RDMA over Converged Ethernet version 2) standard enables RDMA on widely deployed Ethernet technology. RoCEv2 leverages priority-based flow control (PFC) for realizing the lossless environment required by RDMA. However, PFC is well-known to have the technical weakness of head-of-line blocking. Congestion control for RDMA is a very hot research topic for datacenter networks. In this paper, we propose a novel congestion control algorithm for RoCEv2, TIDD (Timer-based Increase and Delay-based Decrease). TIDD basically combines the timer-based increase of DCQCN and delay-based decrease of TIMELY. Extensive simulation results show that TIDD satisfies the high throughput and low latency required for datacenter networks.},
keywords={},
doi={10.1587/transcom.2021WWP0005},
ISSN={1745-1345},
month={April},}
Copy
TY - JOUR
TI - Timer-Based Increase and Delay-Based Decrease Algorithm for RDMA Congestion Control
T2 - IEICE TRANSACTIONS on Communications
SP - 421
EP - 431
AU - Masahiro NOGUCHI
AU - Daisuke SUGAHARA
AU - Miki YAMAMOTO
PY - 2022
DO - 10.1587/transcom.2021WWP0005
JO - IEICE TRANSACTIONS on Communications
SN - 1745-1345
VL - E105-B
IS - 4
JA - IEICE TRANSACTIONS on Communications
Y1 - April 2022
AB - For recent datacenter networks, RDMA (Remote Direct Memory Access) can ease the overhead of the TCP/IP protocol suite. The RoCEv2 (RDMA over Converged Ethernet version 2) standard enables RDMA on widely deployed Ethernet technology. RoCEv2 leverages priority-based flow control (PFC) for realizing the lossless environment required by RDMA. However, PFC is well-known to have the technical weakness of head-of-line blocking. Congestion control for RDMA is a very hot research topic for datacenter networks. In this paper, we propose a novel congestion control algorithm for RoCEv2, TIDD (Timer-based Increase and Delay-based Decrease). TIDD basically combines the timer-based increase of DCQCN and delay-based decrease of TIMELY. Extensive simulation results show that TIDD satisfies the high throughput and low latency required for datacenter networks.
ER -