Search

The emergence of Large Language Model(LLM) technologies has led to a rapidly growing demand for compute resources in models. In response, the enterprises are building large-scale multi-tenant GPU clusters with 10k or even ore GPUs. In contrast to the rapidly growing cluster size, the bandwidth of clusters has also been increasing to meet communication demands, with 800 Gbps optical modules already in practical use and 1.6 Tbps modules on the horizon. However, designing clusters that simultaneously meet the requirements of large scale and high bandwidth is challenging due to the limited capacity of electrical switch chips. Unlike electrical switch chips, the single-port bandwidth of MEMS-OCS is solely determined by the optical module, making it straightforward to achieve both bandwidth and scability requirement. In this paper, we propose an opto-electronic hybrid architecture called \textbf{LumosCore}. We address the issues of L2 protocols incompatibility potential network contention and algorithm time complexity through physical topology and logical topology design. Additionally, we design a polynomial-time complexity link reconfiguration algorithm to reconfigure MEMS-OCS with minimal time overhead. We validate the feasibility of the proposed scheme in a cluster consisting of 128 NPUs, and through simulation based on real traces, we demonstrate the superiority of \textbf{LumosCore} over traditional architectures.

Balancing training on long-tail data distributions remains a long-standing challenge in deep learning. While methods such as re-weighting and re-sampling help alleviate the imbalance issue, limited sample diversity continues to hinder models from learning robust and generalizable feature representations, particularly for tail classes. In contrast to existing methods, we offer a novel perspective on long-tail learning, inspired by an observation: datasets with finer granularity tend to be less affected by data imbalance. In this paper, we investigate this phenomenon through both quantitative and qualitative studies, showing that increased granularity enhances the generalization of learned features in tail categories. Motivated by these findings, we propose a method to increase dataset granularity through category extrapolation. Specifically, we introduce open-set auxiliary classes that are visually similar to existing ones, aiming to enhance representation learning for both head and tail classes. This forms the core contribution and insight of our approach. To automate the curation of auxiliary data, we leverage large language models (LLMs) as knowledge bases to search for auxiliary categories and retrieve relevant images through web crawling. To prevent the overwhelming presence of auxiliary classes from disrupting training, we introduce a neighbor-silencing loss that encourages the model to focus on class discrimination within the target dataset. During inference, the classifier weights for auxiliary categories are masked out, leaving only the target class weights for use. Extensive experiments and ablation studies on three standard long-tail benchmarks demonstrate the effectiveness of our approach, notably outperforming strong baseline methods that use the same amount of data. The code will be made publicly available.

Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage., Comment: 19 pages, 4 figures

Recent advances in fast optical switching technology show promise in meeting the high goodput and low latency requirements of datacenter networks (DCN). We present NegotiaToR, a simple network architecture for optical reconfigurable DCNs that utilizes on-demand scheduling to handle dynamic traffic. In NegotiaToR, racks exchange scheduling messages through an in-band control plane and distributedly calculate non-conflicting paths from binary traffic demand information. Optimized for incasts, it also provides opportunities to bypass scheduling delays. NegotiaToR is compatible with prevalent flat topologies, and is tailored towards a minimalist design for on-demand reconfigurable DCNs, enhancing practicality. Through large-scale simulations, we show that NegotiaToR achieves both small mice flow completion time (FCT) and high goodput on two representative flat topologies, especially under heavy loads. Particularly, the FCT of mice flows is one to two orders of magnitude better than the state-of-the-art traffic-oblivious reconfigurable DCN design., Comment: This paper is accepted by ACM SIGCOMM 2024

Traffic Engineering (TE) is critical for improving network performance and reliability. A key challenge in TE is the management of sudden traffic bursts. Existing TE schemes either do not handle traffic bursts or uniformly guard against traffic bursts, thereby facing difficulties in achieving a balance between normal-case performance and burst-case performance. To address this issue, we introduce FIGRET, a Fine-Grained Robustness-Enhanced TE scheme. FIGRET offers a novel approach to TE by providing varying levels of robustness enhancements, customized according to the distinct traffic characteristics of various source-destination pairs. By leveraging a burst-aware loss function and deep learning techniques, FIGRET is capable of generating high-quality TE solutions efficiently. Our evaluations of real-world production networks, including Wide Area Networks and data centers, demonstrate that FIGRET significantly outperforms existing TE schemes. Compared to the TE scheme currently deployed in Jupiter data center networks of Google, FIGRET achieves a 9\%-34\% reduction in average Maximum Link Utilization and improves solution speed by $35\times$-$1800 \times$. Against DOTE, a state-of-the-art deep learning-based TE method, FIGRET substantially lowers the occurrence of significant congestion events triggered by traffic bursts by 41\%-53.9\% in topologies with high traffic dynamics.

Distributed machine learning (DML) technology makes it possible to train large neural networks in a reasonable amount of time. Meanwhile, as the computing power grows much faster than network capacity, network communication has gradually become the bottleneck of DML. Current multi-tenant GPU clusters face network contention caused by hash-collision problem which not only further increases the overhead of communication, but also creates unfairness and affects the user experience. In this paper, we firstly analyse how network contention affects the training time in a cluster with 32 NVIDIA V100 GPUs. Then we propose vClos to eliminate network contention by jointly optimizing network topology and communication pattern in distributed training. An OCS-vClos which introduces a layer of optical circuit switches (OCSs) in the leaf-spine network is also proposed to reduce potential network resource fragmentation caused by resource allocation strategy in vClos. Testbed experiments and real-trace-based large-scale simulations are conducted to demonstrate the superiority of vClos over existing network resource scheduling strategies.

We study how to reduce the reconfiguration time in hybrid optical-electrical Datacenter Networks (DCNs). With a layer of Optical Circuit Switches (OCSes), hybrid optical-electrical DCNs can reconfigure its logical topologies to better match the on-going traffic patterns, but the reconfiguration time may directly affect the benefits of reconfigurability. The reconfiguration time consists of the topology solver running time and the network convergence time after triggering reconfiguration. However, existing topology solvers either incur high algorithmic complexity or fail to minimize the reconfiguration overhead. In this paper, we propose a novel algorithm that combines the ideas of bipartition and Minimum Cost Flow (MCF) to reduce the overall reconfiguration time. For the first time, we formulate the topology solving problem as a MCF problem with piecewise-linear cost, which strikes a better balance between solver complexity and solution optimality. Our evaluation shows that our algorithm can significantly reduce the network convergence time while consuming less topology solver running time, making its overall performance superior to existing algorithms. Our code and test cases are available at a public repository.

Recent years have witnessed a widespread adoption of containers. While containers simplify and accelerate application development, existing container network technologies either incur significant overhead, which hurts performance for distributed applications, or lose flexibility or compatibility, which hinders the widespread deployment in production. We carefully analyze the kernel data path of an overlay network, quantifying the time consumed by each segment of the data path and identifying the \emph{extra overhead} in an overlay network compared to bare metal. We observe that this extra overhead generates repetitive results among packets, which inspires us to introduce caches within an overlay network. We design and implement ONCache (\textbf{O}verlay \textbf{N}etwork \textbf{Cache}), a cache-based container overlay network, to eliminate the extra overhead while maintaining flexibility and compatibility. We implement ONCache using the extended Berkeley Packet Filter (eBPF) with only 524 lines of code, and integrate it as a plugin of Antrea. With ONCache, containers attain networking performance akin to that of bare metal. Compared to the standard overlay networks, ONCache improves throughput and request-response transaction rate by 12\% and 36\% for TCP (20\% and 34\% for UDP), respectively, while significantly reducing per-packet CPU overhead. Popular distributed applications also benefit from ONCache.

Deterministic Networking (DetNet) is a rising technology that offers deterministic delay \& jitter and zero packet loss regardless of failures in large IP networks. In order to support DetNet, we must be able to find a set of low-cost routing paths for a given node pair subject to delay-range constraints. Unfortunately, the \textbf{Delay-Range} Constrained Routing (DRCR) problem is NP-Complete. Existing routing approaches either cannot support the delay-range constraints, or incur extremely high computational complexity. We propose Pulse$+$, a highly scalable and efficient DRCR problem solver. Pulse$+$ adopts a branch-and-bound methodology and optimizes its pruning strategies for higher efficiency. We also integrate Pulse$+$ with a divide-and-conquer approach and propose CoSE-Pulse$+$ to find a pair of active/backup paths that meet DetNet's delay-range and delay-diff constraints. Both Pulse$+$ and CoSE-Pulse$+$ offer optimality guarantee. Notably, although Pulse$+$ and CoSE-Pulse$+$ do not have a polynomial worst-case time complexity, their empirical performance is superior. We evaluate Pulse$+$ and CoSE-Pulse$+$ against the K-Shorst-Path and Lagrangian-dual based algorithms using synthetic test cases generated over networks with thousands of nodes and links. Both Pulse$+$ and CoSE-Pulse$+$ achieve significant speedup. To enable reproduction, we open source our code and test cases at [1].

Most existing 3D point cloud object detection approaches heavily rely on large amounts of labeled training data. However, the labeling process is costly and time-consuming. This paper considers few-shot 3D point cloud object detection, where only a few annotated samples of novel classes are needed with abundant samples of base classes. To this end, we propose Prototypical VoteNet to recognize and localize novel instances, which incorporates two new modules: Prototypical Vote Module (PVM) and Prototypical Head Module (PHM). Specifically, as the 3D basic geometric structures can be shared among categories, PVM is designed to leverage class-agnostic geometric prototypes, which are learned from base classes, to refine local features of novel categories.Then PHM is proposed to utilize class prototypes to enhance the global feature of each object, facilitating subsequent object localization and classification, which is trained by the episodic training strategy. To evaluate the model in this new setting, we contribute two new benchmark datasets, FS-ScanNet and FS-SUNRGBD. We conduct extensive experiments to demonstrate the effectiveness of Prototypical VoteNet, and our proposed method shows significant and consistent improvements compared to baselines on two benchmark datasets., Comment: NeurIPS 2022

We study how to design edge server placement and server scheduling policies under workload uncertainty for 5G networks. We introduce a new metric called resource pooling factor to handle unexpected workload bursts. Maximizing this metric offers a strong enhancement on top of robust optimization against workload uncertainty. Using both real traces and synthetic traces, we show that the proposed server placement and server scheduling policies not only demonstrate better robustness against workload uncertainty than existing approaches, but also significantly reduce the cost of service providers. Specifically, in order to achieve close-to-zero workload rejection rate, the proposed server placement policy reduces the number of required edge servers by about 25% compared with the state-of-the-art approach; the proposed server scheduling policy reduces the energy consumption of edge servers by about 13% without causing much impact on the service quality., Comment: The extended version of (IWQoS 2021) accepted paper: Design of Robust and Efficient Edge Server Placement and Server Scheduling Policies

Many optical circuit switched data center networks (DCN) have been proposed in the past to attain higher capacity and topology reconfigurability, though commercial adoption of these architectures have been minimal. One major challenge these architectures face is the difficulty of handling uncertain traffic demands using commercial optical circuit switches (OCS) with high switching latency. Prior works have generally focused on developing fast-switching OCS prototypes to quickly react to traffic changes through frequent reconfigurations. This approach, however, adds tremendous complexity to the control plane, and raises the barrier for commercial adoption of optical circuit switched data center networks. We propose COUDER, a robust topology and routing optimization framework for reconfigurable optical circuit switched data centers. COUDER optimizes topology and routing based on a convex set of traffic matrices, and offers strict throughput guarantees for any future traffic matrices bounded by the convex set. For the bursty traffic demands that are unbounded by the convex set, we employ a desensitization technique to reduce performance hit. This enables COUDER to generate topology and routing solutions capable of handling unexpected traffic changes without relying on frequent topology reconfigurations. Our extensive evaluations based on Facebook's production DCN traces show that, even with daily reconfiguration, COUDER achieves about 20\% higher throughput, and about 32\% lower average hop count compared to cost-equivalent static topologies. Our work shows that adoption of reconfigurable topologies in commercial DCNs is feasible even without fast OCSs.

Although Person Re-Identification has made impressive progress, difficult cases like occlusion, change of view-pointand similar clothing still bring great challenges. Besides overall visual features, matching and comparing detailed information is also essential for tackling these challenges. This paper proposes two key recognition patterns to better utilize the detail information of pedestrian images, that most of the existing methods are unable to satisfy. Firstly, Visual Clue Alignment requires the model to select and align decisive regions pairs from two images for pair-wise comparison, while existing methods only align regions with predefined rules like high feature similarity or same semantic labels. Secondly, the Conditional Feature Embedding requires the overall feature of a query image to be dynamically adjusted based on the gallery image it matches, while most of the existing methods ignore the reference images. By introducing novel techniques including correspondence attention module and discrepancy-based GCN, we propose an end-to-end ReID method that integrates both patterns into a unified framework, called CACE-Net((C)lue(A)lignment and (C)onditional (E)mbedding). The experiments show that CACE-Net achieves state-of-the-art performance on three public datasets.

In the conventional person Re-ID setting, it is widely assumed that cropped person images are for each individual. However, in a crowded scene, off-shelf-detectors may generate bounding boxes involving multiple people, where the large proportion of background pedestrians or human occlusion exists. The representation extracted from such cropped images, which contain both the target and the interference pedestrians, might include distractive information. This will lead to wrong retrieval results. To address this problem, this paper presents a novel deep network termed Pedestrian-Interference Suppression Network (PISNet). PISNet leverages a Query-Guided Attention Block (QGAB) to enhance the feature of the target in the gallery, under the guidance of the query. Furthermore, the involving Guidance Reversed Attention Module and the Multi-Person Separation Loss promote QGAB to suppress the interference of other pedestrians. Our method is evaluated on two new pedestrian-interference datasets and the results show that the proposed method performs favorably against existing Re-ID methods., Comment: Accepted by ECCV 2020

Numerous optical circuit switched data center networks have been proposed over the past decade for higher capacity, though commercial adoption of these architectures have been minimal so far. One major challenge commonly facing these architectures is the difficulty of handling bursty traffic with optical circuit switches (OCS) with high switching latency. Prior works generally rely on fast-switching OCS prototypes to better react to traffic changes via frequent reconfigurations. This approach, unfortunately, adds further complexity to the control plane. We propose METTEOR, an easily deployable solution for optical circuit switched data centers, that is designed for the current capabilities of commercial OCSs. Using multiple predicted traffic matrices, METTEOR designs data center topologies that are less sensitive to traffic changes, thus eliminating the need of frequently reconfiguring OCSs upon traffic changes. Results based on extensive evaluations using production traces show that METTEOR increases the percentage of direct-hop traffic by about 80% over a fat tree at comparable cost, and by about 30% over a uniform mesh, at comparable maximum link utilizations. Compared to ideal solutions that reconfigure OCSs on every traffic matrix, METTEOR achieves close-to-optimal bandwidth utilization even with biweekly reconfiguration. This drastically lowers the controller and management complexity needed to perform METTEOR in commercial settings.

We propose a Generative Transfer Network (GTNet) for zero shot object detection (ZSD). GTNet consists of an Object Detection Module and a Knowledge Transfer Module. The Object Detection Module can learn large-scale seen domain knowledge. The Knowledge Transfer Module leverages a feature synthesizer to generate unseen class features, which are applied to train a new classification layer for the Object Detection Module. In order to synthesize features for each unseen class with both the intra-class variance and the IoU variance, we design an IoU-Aware Generative Adversarial Network (IoUGAN) as the feature synthesizer, which can be easily integrated into GTNet. Specifically, IoUGAN consists of three unit models: Class Feature Generating Unit (CFU), Foreground Feature Generating Unit (FFU), and Background Feature Generating Unit (BFU). CFU generates unseen features with the intra-class variance conditioned on the class semantic embeddings. FFU and BFU add the IoU variance to the results of CFU, yielding class-specific foreground and background features, respectively. We evaluate our method on three public datasets and the results demonstrate that our method performs favorably against the state-of-the-art ZSD approaches., Comment: Accepted by AAAI 2020