TY - GEN
T1 - Reverie
T2 - 21st USENIX Symposium on Networked Systems Design and Implementation, NSDI 2024
AU - Addanki, Vamsi
AU - Bai, Wei
AU - Schmid, Stefan
AU - Apostolaki, Maria
N1 - Publisher Copyright:
© 2024 Proceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation, NSDI 2024. All rights reserved.
PY - 2024
Y1 - 2024
N2 - The switch buffers in datacenters today are dynamically shared by traffic classes with different loss tolerance and reaction to congestion signals. In particular, while legacy applications use loss-tolerant transport, e.g., DCTCP, newer applications require lossless datacenter transport, e.g., RDMA over Converged Ethernet. Unfortunately, as we analytically show in this paper, the buffer-sharing practices of today’s datacenters pose a fundamental limitation to effectively isolate RDMA and TCP while also maximizing burst absorption. We identify two root causes: (i) the buffer-sharing for RDMA and TCP relies on two independent and often conflicting views of the buffer, namely ingress and egress; and (ii) the buffer-sharing scheme micromanages the buffer and overreacts to the changes in its occupancy during transient congestion. In this paper, we present REVERIE, a buffer-sharing scheme, which, unlike prior works, is suitable for both lossless and loss-tolerant traffic, providing isolation and better burst absorption than state-of-the-art buffer-sharing schemes. At the core of REVERIE lies a unified (consolidated ingress and egress) admission control that jointly optimizes the buffers for both RDMA and TCP. REVERIE allocates buffer based on a low-pass filter that naturally absorbs bursty queue lengths during transient congestion within the buffer limits. Our evaluation shows that REVERIE can improve the performance of RDMA as well as TCP in terms of flow completion times by up to 33%.
AB - The switch buffers in datacenters today are dynamically shared by traffic classes with different loss tolerance and reaction to congestion signals. In particular, while legacy applications use loss-tolerant transport, e.g., DCTCP, newer applications require lossless datacenter transport, e.g., RDMA over Converged Ethernet. Unfortunately, as we analytically show in this paper, the buffer-sharing practices of today’s datacenters pose a fundamental limitation to effectively isolate RDMA and TCP while also maximizing burst absorption. We identify two root causes: (i) the buffer-sharing for RDMA and TCP relies on two independent and often conflicting views of the buffer, namely ingress and egress; and (ii) the buffer-sharing scheme micromanages the buffer and overreacts to the changes in its occupancy during transient congestion. In this paper, we present REVERIE, a buffer-sharing scheme, which, unlike prior works, is suitable for both lossless and loss-tolerant traffic, providing isolation and better burst absorption than state-of-the-art buffer-sharing schemes. At the core of REVERIE lies a unified (consolidated ingress and egress) admission control that jointly optimizes the buffers for both RDMA and TCP. REVERIE allocates buffer based on a low-pass filter that naturally absorbs bursty queue lengths during transient congestion within the buffer limits. Our evaluation shows that REVERIE can improve the performance of RDMA as well as TCP in terms of flow completion times by up to 33%.
UR - https://www.scopus.com/pages/publications/85182566012
UR - https://www.scopus.com/pages/publications/85182566012#tab=citedBy
M3 - Conference contribution
AN - SCOPUS:85182566012
T3 - Proceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation, NSDI 2024
SP - 651
EP - 668
BT - Proceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation, NSDI 2024
PB - USENIX Association
Y2 - 16 April 2024 through 18 April 2024
ER -