An official website of the United States government
With the ever-increasing size of training models and datasets, network communication has emerged as a major bottleneck in distributed deep learning training. To address this challenge, we propose an optical distributed deep learning (ODDL) architecture. ODDL utilizes a fast yet scalable all-optical network architecture to accelerate distributed training. One of the key features of the architecture is its flow-based transmit scheduling with fast reconfiguration. This allows ODDL to allocate dedicated optical paths for each traffic stream dynamically, resulting in low network latency and high network utilization. Additionally, ODDL provides physically isolated and tailored network resources for training tasks by reconfiguring the optical switch using LCoS-WSS technology. The ODDL topology also uses tunable transceivers to adapt to time-varying traffic patterns. To achieve accurate and fine-grained scheduling of optical circuits, we propose an efficient distributed control scheme that incurs minimal delay overhead. Our evaluation on real-world traces showcases ODDL’s remarkable performance. When implemented with 1024 nodes and 100 Gbps bandwidth, ODDL accelerates VGG19 training by 1.6× and 1.7× compared to conventional fat-tree electrical networks and photonic SiP-Ring architectures, respectively. We further build a four-node testbed, and our experiments show that ODDL can achieve comparable training time compared to that of anidealelectrical switching network.
more »
« less
Scalable Tail Latency Estimation for Data Center Networks
In this paper, we consider how to provide fast estimates of flow-level tail latency performance for very large scale data center networks. Network tail latency is often a crucial metric for cloud application performance that can be affected by a wide variety of factors, including network load, inter-rack traffic skew, traffic burstiness, flow size distributions, oversubscription, and topology asymmetry. Network simulators such as ns-3 and OMNeT++ can provide accurate answers, but are very hard to parallelize, taking hours or days to answer what if questions for a single configuration at even moderate scale. Recent work with MimicNet has shown how to use machine learning to improve simulation performance, but at a cost of including a long training step per configuration, and with assumptions about workload and topology uniformity that typically do not hold in practice. We address this gap by developing a set of techniques to provide fast performance estimates for large scale networks with general traffic matrices and topologies. A key step is to decompose the problem into a large number of parallel independent single-link simulations; we carefully combine these link-level simulations to produce accurate estimates of end-to-end flow level performance distributions for the entire network. Like MimicNet, we exploit symmetry where possible to gain additional speedups, but without relying on machine learning, so there is no training delay. On a large-scale net- work where ns-3 takes 11 to 27 hours to simulate five seconds of network behavior, our techniques run in one to two minutes with accuracy within 9% for tail flow completion times.
more »
« less
- Award ID(s):
- 2006346
- PAR ID:
- 10495839
- Publisher / Repository:
- USENIX Association
- Date Published:
- Journal Name:
- 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23)
- ISBN:
- 978-1-939133-33-5
- Format(s):
- Medium: X
- Location:
- Boston, MA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Modern end-host network stacks have to handle traffic from tens of thousands of flows and hundreds of virtual machines per single host, to keep up with the scale of modern clouds. This can cause congestion for traffic egressing from the end host. The effects of this congestion have received little attention. Currently, an overflowing queue, like a kernel queuing discipline, will drop incoming packets. Packet drops lead to worse network and CPU performance by inflating the time to transmit the packet as well as spending extra effort on retransmissions. In this paper, we show that current end-host mechanisms can lead to high CPU utilization, high tail latency, and low throughput in cases of congestion of egress traffic within the end host. We present zD, a framework for applying backpressure from a congested queue to traffic sources at end hosts that can scale to thousands of flows. We implement zD to apply backpressure in two settings: i) between TCP sources and kernel queuing discipline, and ii) between VMs as traffic sources and kernel queuing discipline in the hypervisor. zD improves throughput by up to 60%, and improves tail RTT by at least 10x at high loads, compared to standard kernel implementation.more » « less
-
Virtual switches, used for end-host networking, drop packets when the receiving application is not fast enough to consume them. This is called the slow receiver problem, and it is important because packet loss hurts tail communication latency and wastes CPU cycles, resulting in application-level performance degradation. Further, solving this problem is challenging because application throughput is highly variable over short timescales as it depends on workload, memory contention, and OS thread scheduling. This paper presents Backdraft, a new lossless virtual switch that addresses the slow receiver problem by combining three new components: (1) Dynamic Per-Flow Queuing (DPFQ) to prevent HOL blocking and provide on-demand memory usage; (2) Doorbell queues to reduce CPU overheads; (3) A new overlay network to avoid congestion spreading. We implemented Backdraft on top of BESS and conducted experiments with real applications on a 100 Gbps cluster with both DCTCP and Homa, a state-of-the-art congestion control scheme. We show that an application with Backdraft can achieve up to 20x lower tail latency at the 99th percentile.more » « less
-
Scale-out datacenter network fabrics enable network operators to translate improved link and switch speeds directly into end-host throughput. Unfortunately, limits in the underlying CMOS packet switch chip manufacturing roadmap mean that NICs, links, and switches are not getting faster fast enough to meet demand. As a result, operators have introduced alternative, parallel fabric designs in the core of the network that deliver N-times the bandwidth by simply forwarding traffic over any of N parallel network fabrics. In this work, we consider extending this parallel network idea all the way to the end host. Our initial impressions found that direct application of existing path selection and forwarding techniques resulted in poor performance. Instead, we show that appropriate path selection and forwarding protocols can not only improve the performance of existing, homogeneous parallel fabrics, but enable the development of heterogeneous parallel network fabrics that can deliver even higher bandwidth, lower latency, and improved resiliency than traditional designs constructed from the same constituent components.more » « less
-
5G New Radio cellular networks are designed to provide high Quality of Service for application on wirelessly connected devices. However, changing conditions of the wireless last hop can degrade application performance, and the applications have no visibility into the 5G Radio Access Network (RAN). Most 5G network operators run closed networks, limiting the potential for co-design with the wider-area internet and user applications. This paper demonstrates NR-Scope, a passive, incrementally-deployable, and independently-deployable Standalone 5G network telemetry system that can passively measure fine-grained RAN capacity, latency, and retransmission information. Application servers can take advantage of the measurements to achieve better millisecond scale, application-level decisions on offered load and bit rate adaptation than end-to-end latency measurements or end-to-end packet losses currently permit. We demonstrate the performance of NR-Scope by decoding the downlink control information (DCI) for downlink and uplink traffic of a 5G Standalone base station in real-time.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
