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 an
This content will become publicly available on November 28, 2024
- Award ID(s):
- 1662487
- NSF-PAR ID:
- 10483618
- Publisher / Repository:
- ACM
- Date Published:
- Page Range / eLocation ID:
- 131 to 138
- Format(s):
- Medium: X
- Location:
- Cambridge MA USA
- Sponsoring Org:
- National Science Foundation
More Like this
-
more » « less
ideal electrical switching network. -
Packet drops caused by congestion are a fundamental problem in network operation. Yet, it is difficult to detect where drops are happening, let alone which flows are most affected. Detecting the small-timescale drops caused by short bursts of traffic is even more challenging, and traditional monitoring techniques can easily miss them. To uncover packet drops as they occur inside a switch, the analysis must be real-time, fine-grained, and efficient. However, modern switches have distributed packet-processing pipelines that see either the arriving or departing traffic, but not the packet drops. Plus, they do not have enough memory to store per-flow state. Our MIDST system addresses these challenges through a distributed compact data structure with lightweight coordination between ingress and egress pipelines. MIDST identifies the flows experiencing loss, as well as the bursty flows responsible, across different burst durations. Our evaluation with real-world traces and TCP connections shows that MIDST uses little memory (e.g., 320KB) while providing high accuracy (95% to 98%) under varying loss rates and burst durations. We evaluate a low-rate DDoS attack and demonstrate the potential use of our measurement results for attack detection and mitigation.more » « less
-
Activity recognition has applications in a variety of human-in-the-loop settings such as smart home health monitoring, green building energy and occupancy management, intelligent transportation, and participatory sensing. While fine-grained activity recognition systems and approaches help enable a multitude of novel applications, discovering them with non-intrusive ambient sensor systems pose challenging design, as well as data processing, mining, and activity recognition issues. In this paper, we develop a low-cost heterogeneous Radar based Activity Monitoring (RAM) system for recognizing fine-grained activities. We exploit the feasibility of using an array of heterogeneous micro-doppler radars to recognize low-level activities. We prototype a short-range and a long-range radar system and evaluate the feasibility of using the system for fine-grained activity recognition. In our evaluation, using real data traces, we show that our system can detect fine-grained user activities with 92.84% accuracy.more » « less
-
The growing number of network functions drives the need to install increasing numbers of fine-grained packet classification rules in the network switches. However, this demand for rules is outstripping the size of switch memory. While much work has focused on compressing classification rules, most of this work proposes solutions operating in the memory of a single switch. This paper proposed, instead, a collaborative approach encompassing switches and network functions: we couple approximate classification at switches with fine-grained filtering where needed at network functions to accomplish overall classification. This architecture enables a trade-off between usage of (expensive) switch memory and (cheaper) downstream network bandwidth and network function resources. Our implementation uses approximate classification and Prefiltering to reduce switch memory usage. Our system can reduce memory significantly compared to a strawman approach, as shown by simulations of real traffic traces and rules.more » « less
-
null (Ed.)more » « less
This paper introduces a hierarchical traffic model for spread measurement of network traffic flows. The hierarchical model, which aggregates lower level flows into higher-level flows in a hierarchical structure, will allow us to measure network traffic at different granularities at once to support diverse traffic analysis from a grand view to fine-grained details. The spread of a flow is the number of distinct elements (under measurement) in the flow, where the flow label (that identifies packets belonging to the flow) and the elements (which are defined based on application need) can be found in packet headers or payload. Traditional flow spread estimators are designed without hierarchical traffic modeling in mind, and incur high overhead when they are applied to each level of the traffic hierarchy. In this paper, we propose a new Hierarchical Virtual bitmap Estimator (HVE) that performs simultaneous multi-level traffic measurement, at the same cost of a traditional estimator, without degrading measurement accuracy. We implement the proposed solution and perform experiments based on real traffic traces. The experimental results demonstrate that HVE improves measurement throughput by 43% to 155%, thanks to the reduction of perpacket processing overhead. For small to medium flows, its measurement accuracy is largely similar to traditional estimators that work at one level at a time. For large aggregate and base flows, its accuracy is better, with up to 97% smaller error in our experiments.
