Less Is More: Trading a Little Bandwidth for Ultra-Low Latency in the Data Center

By Mohammad Alizadeh, A. Kabbani, Tom Edsall, Balaji Prabhakar, Amin Vahdat, Masato Yasuda. Symposium on Networked Systems Design and Implementation, 2012

503 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/61aa09bc1a3eae17480645c90b06a18cbd62d9af

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Reproducible network experiments using container-based emulation (2012, background): “HULL [38] By sacriﬁcing a small amount of bandwidth, HULL can reduce average and tail latencies in data center networks. replicated MPTCP [39] Multipath TCP increases performance over multiple wireless interfaces versus TCP and can perform seamless wireless handoffs. replicated Over 3G and WiFi,…”
Fastpass (2014, background): “DCTCP[6] and HULL [7] reduce switch queueing, but increase convergence time to a fair-share of the capacity, and do not eliminate queueing delay.”
Copa: Practical Delay-Based Congestion Control for the Internet (2018, background): “The past few years have seen new congestion control methods for datacenters [1, 2, 3, 36], cellular networks [44, 46], Web applications [12], video streaming [13, 25], vehicular Wi-Fi [11, 26], and more.”
Ultra-Low Latency (ULL) Networks: The IEEE TSN and IETF DetNet Standards and Related 5G ULL Research (2018, background): “Latency reduction techniques for data center networks have received increasing attention in recent years, see for instance [376]–[406].”
Homa: a receiver-driven low-latency transport protocol using network priorities (2018, methodology): “ute performance (14 s round-trip for small messages at 80% network load and99thpercentiletaillatency)isnearlytwoordersofmagnitude faster than the best available comparison systems. For example, HULL [3] reported 782 s one-way latency for 1 Kbyte messages at 99th percentile and 60% network load, and ”
pFabric: minimal near-optimal datacenter transport (2013, methodology): “Implicit rate control: DCTCP [3] and HULL [4] try to keep the fabric queues small or empty by employing an adaptive congestion control algorithm based on ECN and other mechanisms such as operating the network at slightly less than 100% utilization, packet pacing, etc to appropriately throttle long…”
TIMELY: RTT-based Congestion Control for the Datacenter (2015, methodology): “Existing datacenter transports use signals from network switches to detect the onset of congestion and run with low levels of latency and loss [15,35,37].”
DeTail: reducing the flow completion time tail in datacenter networks (2012, background): “4 Current Solutions Insufficient DCTCP, D, and HULL [12,13,33] are single path solutions recently proposed to reduce the completion times of latency-sensitive flows.”
Presto: Edge-based Load Balancing for Fast Datacenter Networks (2015, methodology): “HULL [6] uses Phantom Queues and congestion notiﬁca-tions to cap link utilization and prevent congestion.”
Congestion Control for Large-Scale RDMA Deployments (2015, methodology): “Instead, like prior work [2, 5, 6], we extract salient characteristics of the trace data, such as ﬂow size distribution and generate synthetic trafﬁc to match these characteristics.”
PINT: Probabilistic In-band Network Telemetry (2020, background): “This means that even a small amount of bandwidth headroom can provide a dramatic reduction in latency [4].”
HPCC: high precision congestion control (2019, background): “The slowdown increases with a higher load as the theory in [9] shows.”
Swift: Delay is Simple and Effective for Congestion Control in the Datacenter (2020, methodology): “Hull [2] uses phantom queues, and D2TCP [52] is a deadline-aware protocol that also uses ECN.”
On the impact of packet spraying in data center networks (2013, methodology): “Solutions such as DCTCP [10] and HULL [11] provide even lower latencies at the (slight) expense of throughput.”
Re-architecting datacenter networks and stacks for low latency and high performance (2017, background): “Hull [5] reserves capacity to ensure low latency for short flows and uses phantom queues to identify potential congestion; it requires accurate packet pacing at endpoints.”