Interactive notebook programming is universal in modern ML and AI workflows, with interactive deep learning training (IDLT) emerging as a dominant use case.To ensure responsiveness, platforms like Jupyter and Colab reserve GPUs for long-running notebook sessions, despite their intermittent and sporadic GPU usage, leading to extremely low GPU utilization and prohibitively high costs.In this paper, we introduce NotebookOS, a GPU-efficient notebook platform tailored for the unique requirements of IDLT.NotebookOS employs replicated notebook kernels with Raft-synchronized replicas distributed across GPU servers.To optimize GPU utilization, NotebookOS oversubscribes server resources, leveraging high inter-arrival times in IDLT workloads, and allocates GPUs only during active cell execution.It also supports replica migration and automatic cluster scaling under high load.Altogether, this design enables interactive training with minimal delay.In evaluation on production workloads, NotebookOS saved over 1,187 GPU hours in 17.5 hours of real-world IDLT, while significantly improving interactivity.

Leveraging graph structures for time-series forecasting has garnered significant attention due to their effective relationship modeling between nodes and their associated time-series. However, in scenarios entities communicate in a broadcasting manner, graph models fall short of pairwise modeling. Hypergraph models address this by capturing beyond-pairwise interactions among node time-series. Nevertheless, most hypergraph models overlook the dynamics between nodes and their incident hyperedges, assuming constant node-hyperedge connections. In this paper, we introduce a novel model, Probabilistic Hypergraph Recurrent Neural Networks (PHRNN), which leverages node-hyperedge dynamics for accurate time-series forecasting. PHRNN associates each time-series with a node and models node interactions on a hypergraph, capturing beyond-pairwise interactions. Moreover, PHRNN learns a probabilistic hypergraph in which node-hyperedge relations are modeled as probabilistic distributions instead of fixed values, capturing dynamic node-hyperedge relations. PHRNN further integrates a prior knowledge KNN hypergraph as regularization when learning the probabilistic hypergraph structure. To the best of our knowledge, PHRNN is the first time-series forecasting model that incorporates hypergraph modeling and probabilistic relationship modeling. Forecasting results from extensive experiments show that PHRNN outperforms state-of-the-art graph and hypergraph baselines on real-world datasets.

Time-series forecasting models are invariably used in a variety of domains for crucial decision-making. Traditionally these models are constructed by experts with considerable manual effort. Unfortunately, this approach has poor scalability while generating accurate forecasts for new datasets belonging to diverse applications. Without access to skilled domain-knowledge, one approach is to train all the models on the new time-series data and then select the best one. However, this approach is nonviable in practice. In this work, we develop techniques for fast automatic selection of the best forecasting model for a new unseen time-series dataset, without having to first train (or evaluate) all the models on the new time-series data to select the best one. In particular, we develop a forecasting meta-learning approach called AutoForecast that allows for the quick inference of the best time-series forecasting model for an unseen dataset. Our approach learns both forecasting models’ performances over time horizon of the same dataset and task similarity across different datasets. The experiments demonstrate the effectiveness of the approach over state-of-the-art (SOTA) single and ensemble methods and several SOTA meta-learners (adapted to our problem) in terms of selecting better forecasting models (i.e., 2 \(\times\) gain) for unseen tasks for univariate and multivariate testbeds. AutoForecast has also significant reduction in inference time compared to the naïve approach (doing inference using all possible models and then selecting the best one), with median of 42 \(\times\) across the two testbeds. We release our meta-learning database corpus (348 datasets), performances of the 322 forecasting models on the database corpus, meta-features, and source codes for the community to access them for forecasting model selection and to build on them with new datasets and models which can help advance automating time-series forecasting problem. In our released database corpus, we unveil new traces of Adobe computing cluster usage for production workloads.

Recent successes of foundation models in large language models have inspired researchers to apply similar technologies to time-series forecasting. Unlike conventional time-series forecasting models, which are trained on the training subset of the target dataset, foundation models are trained on a large collection of source datasets that do not necessarily include the target dataset, with the assumption that foundation models can capture the complex patterns between the input time-series values and the desired predictions. Although many foundation models have claimed superior prediction performance compared to conventional models, one question remains unanswered: Do foundation models for time-series forecasting, which train on many datasets other than the target dataset, perform better than conventional models that train on only (the training subset of) the target dataset? To answer this question, this paper adapts a diffusion-based foundation model and conducts extensive experiments using both small datasets and a large collection of over 100 datasets. Our results show that training on large-scale datasets does not necessarily guarantee a better performance than a conventional model that trains only on the dataset from the same domain. Hence, this paper provides insights for future foundation model research in large-scale time-series forecasting, emphasizing that the usage of target datasets should be considered in addition to training on large-scale source datasets.

With the advent of cloud platforms that offer GPU-as-a-Service (GPUaaS), such as Amazon EC2 and Microsoft Azure, researchers increasingly rely on virtual GPU instances for training deep learning (DL) workloads. These GPU instances vary in configuration attributes, including but not limited to the number of GPUs, the number of vCPUs, and per-hour usage cost. Identifying the appropriate GPU instance for training a DL workload becomes extremely difficult due to the huge GPU instance selection space offered by the cloud platforms and the corresponding variation in training performance or computational needs of different DL workloads. In this paper, we propose a GPU instance recommendation system called GPURank, which provides a recommended list of GPU instances to choose from for DL workloads. GPURank predicts and leverages two metrics: epoch training cost and average GPU utilization to make this choice. We curated a new benchmark dataset by profiling diverse DL workloads to train the regression models in GPURank’s prediction framework. We demonstrate that GPURank beats baselines on two pertinent problem settings: (1) unseen workloads and (2) unseen GPU instances, with a 25.89% and 20.10% higher average ranking performance on these respectively.

Kaige Xie, Tong Yu, Haoliang Wang, Junda Wu, Handong Zhao, Ruiyi Zhang, Kanak Mahadik, Ani Nenkova, Mark Riedl. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Many existing deep graph models have shown that forecasting time-series values benefits from modeling the mutual relations between time-series. For example, graph neural networks can exploit the correlations between two CPU utilization time-series, facilitating more accurate predictions. However, the implied pairwise interactions between entities in the graph structure do not always reflect the actual interactions. In a cloud system, for instance, computing tasks are assigned to groups of machines, and CPU utilization time-series within the same group simultaneously interact with one another. Hence, such interactions are beyond-pairwise. In this paper, we propose a novel model called Hypergraph Recurrent Neural Networks (HGRNN) for time-series forecasting. Our model employs a hypergraph to model beyond-pairwise relations, which naturally reflect the actual interactions among entities. We also introduce a novel semi-principled hypergraph construction method to address the challenge of missing hypergraph information. Our model adopts the encoder-decoder framework where historical time-series are digested into an encoded state which is decoded to yield prediction. We further integrate a temporal component to enhance learning from temporal locality. Extensive experiments on large-scale datasets show that our model achieves better forecasting performance compared to state-of-the-art baselines.

Effective time-series forecasting methods are of significant importance to solve a broad spectrum of research problems. Deep probabilistic forecasting techniques have recently been proposed for modeling large collections of time-series. However, these techniques explicitly assume either complete independence (local model) or complete dependence (global model) between time-series in the collection. This corresponds to the two extreme cases where every time-series is disconnected from every other time-series in the collection or likewise, that every time-series is related to every other time-series resulting in a completely connected graph. In this work, we propose a deep hybrid probabilistic graph-based forecasting framework called Graph Deep Factors (GraphDF) that goes beyond these two extremes by allowing nodes and their time-series to be connected to others in an arbitrary fashion. GraphDF is a hybrid forecasting framework that consists of a relational global and relational local model. In particular, a relational global model learns complex non-linear time-series patterns globally using the structure of the graph to improve both forecasting accuracy and computational efficiency. Similarly, instead of modeling every time-series independently, a relational local model not only considers its individual time-series but also the time-series of nodes that are connected in the graph. The experiments demonstrate the effectiveness of the proposed deep hybrid graph-based forecasting model compared to the state-of-the-art methods in terms of its forecasting accuracy, runtime, and scalability. Our case study reveals that GraphDF can successfully generate cloud usage forecasts and opportunistically schedule workloads to increase cloud cluster utilization by 47.5% on average. Furthermore, we target addressing the common nature of many time-series forecasting applications where time-series are provided in a streaming version; however, most methods fail to leverage the newly incoming time-series values and result in worse performance over time. In this article, we propose an online incremental learning framework for probabilistic forecasting. The framework is theoretically proven to have lower time and space complexity. The framework can be universally applied to many other machine learning-based methods.

Online recommender systems have proven to have ubiquitous applications in various domains. To provide accurate recommendations in real time it is imperative to constantly train and deploy models with the latest data samples. This retraining involves adjusting the model weights by incorporating newly-arrived streaming data into the model to bridge the accuracy gap. To provision resources for the retraining, typically the compute is hosted on VMs, however, due to the dynamic nature of the data arrival patterns, stateless functions would be an ideal alternative over VMs, as they can instantaneously scale on demand. However, it is non-trivial to statically configure the stateless functions because the model retraining exhibits varying resource needs during different phases of retraining. Therefore, it is crucial to dynamically configure the functions to meet the resource requirements, while bridging the accuracy gap. In this paper, we propose Sandpiper, an adaptive framework that leverages stateless functions to deliver accurate predictions at low cost for online recommender systems. The three main ideas in Sandpiper are (i) we design a data-drift monitor that automatically triggers model retraining at required time intervals to bridge the accuracy gap due to incoming data drifts; (ii) we develop an online configuration model that selects the appropriate function configurations while maintaining the model serving accuracy within the latency and cost budget; and (iii) we propose a dynamic synchronization policy for stateless functions to speed up the distributed model retraining leading to cloud cost minimization. A prototype implementation on AWS shows that Sandpiper maintains the average accuracy above 90%, while 3.8× less expensive than the traditional VM-based schemes.

In this work, we develop techniques for fast automatic selection of the best forecasting model for a new unseen time-series dataset, without having to first train (or evaluate) all the models on the new time-series data to select the best one. In particular, we develop a forecasting meta-learning approach called AutoForecast that allows for the quick inference of the best time-series forecasting model for an unseen dataset. Our approach learns both forecasting models performances over time horizon of same dataset and task similarity across different datasets. The experiments demonstrate the effectiveness of the approach over state-of-the-art (SOTA) single and ensemble methods and several SOTA meta-learners (adapted to our problem) in terms of selecting better forecasting models (i.e., 2X gain) for unseen tasks for univariate and multivariate testbeds.

Web applications utilize Recommendation Systems (RS) to address the problem of consumer over-choices. Recent works have taken advantage of multi-modality or multi-view, input information (such as user interaction, images, texts, rating scores) to boost recommendation system performance compared with using single-modality information. However, the use of multi-modality input demands much higher computational cost and storage capacity. On the other hand, the real-world RS services usually have strict budgets on both time and space for a good customer experience. As a result, the model efficiency of multi-modality recommendation systems has gained increasing importance. While unfortunately, to the best of our knowledge, there is no existing study of a generic compression framework for multi-modality RS. In this paper, we investigate, for the first time, how to compress a multi-modality recommendation system with a fixed budget. Assuming that input information from different modalities are of unequal importance, a good compression algorithm should learn to automatically allocate different resource budgets to each input, based on their importance in maximally preserving recommendation efficacy. To this end, we leverage the tools of neural architecture search (NAS) and distillation and propose Auto Multi-modAlity Recommendation System (AutoMARS), a unified modality-aware model compression framework dedicated to multi-modality recommendation systems. We demonstrate the effectiveness and generality of AutoMARS by testing it on three different Amazon datasets of various sparsity. AutoMARS demonstrates superior multi-modality compression performance than previous state-of-the-art compression methods. For example on the Amazon Beauty dataset, we achieve on average a 20% higher accuracy over previous state-of-the-art methods, while enjoying 65% reduction over baselines. Codes are available at: https://github.com/VITA-Group/AutoMARS.

We study the correlated stochastic knapsack problem of a submodular target function, with optional additional constraints. We utilize the multilinear extension of submodular function, and bundle it with an adaptation of the relaxed linear constraints from Ma [Mathematics of Operations Research, Volume 43(3), 2018] on correlated stochastic knapsack problem. The relaxation is then solved by the stochastic continuous greedy algorithm, and rounded by a novel method to fit the contention resolution scheme (Feldman et al. [FOCS 2011]). We obtain a pseudo-polynomial time $(1 - 1/\sqrt{e})/2 \simeq 0.1967$ approximation algorithm with or without those additional constraints, eliminating the need of a key assumption and improving on the $(1 - 1/\sqrt[4]{e})/2 \simeq 0.1106$ approximation by Fukunaga et al. [AAAI 2019].

BACKGROUND: Sequencing technologies are prone to errors, making error correction (EC) necessary for downstream applications. EC tools need to be manually configured for optimal performance. We find that the optimal parameters (e.g., k-mer size) are both tool- and dataset-dependent. Moreover, evaluating the performance (i.e., Alignment-rate or Gain) of a given tool usually relies on a reference genome, but quality reference genomes are not always available. We introduce Lerna for the automated configuration of k-mer-based EC tools. Lerna first creates a language model (LM) of the uncorrected genomic reads, and then, based on this LM, calculates a metric called the perplexity metric to evaluate the corrected reads for different parameter choices. Next, it finds the one that produces the highest alignment rate without using a reference genome. The fundamental intuition of our approach is that the perplexity metric is inversely correlated with the quality of the assembly after error correction. Therefore, Lerna leverages the perplexity metric for automated tuning of k-mer sizes without needing a reference genome. RESULTS: First, we show that the best k-mer value can vary for different datasets, even for the same EC tool. This motivates our design that automates k-mer size selection without using a reference genome. Second, we show the gains of our LM using its component attention-based transformers. We show the model's estimation of the perplexity metric before and after error correction. The lower the perplexity after correction, the better the k-mer size. We also show that the alignment rate and assembly quality computed for the corrected reads are strongly negatively correlated with the perplexity, enabling the automated selection of k-mer values for better error correction, and hence, improved assembly quality. We validate our approach on both short and long reads. Additionally, we show that our attention-based models have significant runtime improvement for the entire pipeline-18[Formula: see text] faster than previous works, due to parallelizing the attention mechanism and the use of JIT compilation for GPU inferencing. CONCLUSION: Lerna improves de novo genome assembly by optimizing EC tools. Our code is made available in a public repository at: https://github.com/icanforce/lerna-genomics .

Traditional resource (demand or supply) forecasting models mainly focus on modeling temporal dependency. However, spatio-temporal data include complex non-linear relational and spatial dependencies. In addition, dynamic contextual information also impacts resources. Methods that consider context assume that the impact of context on resources is fixed, which is not realistic. For example, in a bicycle-sharing system, bike supply in stations is affected by the weather, and that effect changes over time. We propose a novel graph-based context integrated relational model, Context Integrated Graph Neural Network (CIGNN), which models temporal, relational, spatial, and dynamic contextual dependencies for multi-step ahead resource forecasting. We define a resource graph, where nodes represent locations with associated resource time-series, and context graphs (one for each type of context), where nodes represent locations with associated contextual time-series. Assuming that various contexts have dynamic impact on resources, our proposed CIGNN model employs a novel fusion mechanism that jointly learns from multiple contextual time-series. To the best of our knowledge, CIGNN is the first approach that integrates dynamic contextual information using graph neural networks for resource forecasting. Empirical results on two real-world datasets demonstrate that CIGNN consistently outperforms state-of-the-art approaches.

Deep probabilistic forecasting techniques have recently been proposed for modeling large collections of time-series. However, these techniques explicitly assume either complete independence (local model) or complete dependence (global model) between time-series in the collection. This corresponds to the two extreme cases where every time-series is disconnected from every other time-series in the collection or likewise, that every time-series is related to every other time-series resulting in a completely connected graph. In this work, we propose a deep hybrid probabilistic graph-based forecasting framework called Graph Deep Factors (GraphDF) that goes beyond these two extremes by allowing nodes and their time-series to be connected to others in an arbitrary fashion. GraphDF is a hybrid forecasting framework that consists of a relational global and relational local model. In particular, we propose a relational global model that learns complex non-linear time-series patterns globally using the structure of the graph to improve both forecasting accuracy and computational efficiency. Similarly, instead of modeling every time-series independently, we learn a relational local model that not only considers its individual time-series but also the time-series of nodes that are connected in the graph. The experiments demonstrate the effectiveness of the proposed deep hybrid graph-based forecasting model compared to the state-of-the-art methods in terms of its forecasting accuracy, runtime, and scalability. Our case study reveals that GraphDF can successfully generate cloud usage forecasts and opportunistically schedule workloads to increase cloud cluster utilization by 47.5% on average.

Deep neural networks have been widely used in video analysis applications such as automatic metadata generation, action recognition, and video summarization. A fundamental module in the pipeline of such DNN-based applications is feature extraction. However, extracting features for videos is a major bottleneck since it is performed on every frame of each video sequentially. In addition, the long training time of these complex networks also hinders their usability. In this work, we identify fine-grained and coarse-grained parallelism techniques to speed up vital components in video analysis applications through inter-frame and intra-video parallelism. We demonstrate these techniques on the feature extraction and summarization modules. We leverage frame-level parallelism in feature extraction and intra-video parallelism to speed up video summarization and implement them in a distributed environment using Hadoop Map-Reduce framework to get a speed up of 2.67X on a 4-node setup. Furthermore, we show in our results that our approach has similar accuracy to the sequential applications.

Contextual bandit algorithms are commonly used in recommender systems, where content popularity can change rapidly. These algorithms continuously learn latent mappings between users and items, based on contexts associated with them both. Recent recommendation algorithms that learn clustering or social structures between users have exhibited higher recommendation accuracy. However, as the number of users and items in the environment increases, the time required to generate recommendations deteriorates significantly. As a result, these cannot be deployed in practice. The state-of-the-art distributed bandit algorithm - DCCB - relies on a peer-to-peer network to share information among distributed workers. However, this approach does not scale well with the increasing number of users. Furthermore, it suffers from slow discovery of clusters, resulting in accuracy degradation.

No abstract available.

High-throughput next generation sequencers (NGS) can rapidly read billions of short DNA fragments, called reads, at low cost. Moreover, their throughput is increasing and cost is decreasing at rates much faster than the Moore's law. This demands commensurate acceleration for NGS secondary analysis that process the reads to identify variations between genomes. Conventional architectural improvements can at best improve performance at the rate of Moore's law even if the software tools efficiently utilize the underlying architecture. Unfortunately, most of the dozens of software products developed for this purpose fail to exploit the underlying architecture well. Therefore, to match the pace of development of the sequencers, we will need architecture that is more tailored for the computational requirements of NGS secondary analysis as well as software that uses the architecture optimally.

Breakthroughs in gene sequencing technologies have led to an exponential increase in the amount of genomic data. Efficient tools to rapidly process such large quantities of data are critical in the study of gene functions, diseases, evolution, and population variation. These tools are designed in an ad-hoc manner, and require extensive programmer effort to develop and optimize them. Often, such tools are written with the currently available data sizes in mind, and soon start to under perform due to the exponential growth in data. Furthermore, to obtain high-performance, these tools require parallel implementations, adding to the development complexity.

The cost of DNA sequencing has plummeted in recent years. The consequent data deluge has imposed big burdens for data analysis applications. For example, MG-RAST, a production open-public metagenome annotation service, has experienced increasingly large amount of data submission and has demanded scalable resources for the computational needs. To address this problem, we have developed a scalable platform to port MG-RAST workloads into the cloud, where elastic computing resources can be used on demand. To efficiently utilize such resources, however, one must understand the characteristics of the application workloads. In this paper, we characterize the MG-RAST workloads running in the cloud, from the perspectives of computation, I/O, and data transfer. Insights from this work will help guide application enhancement, service operation, and resource management for MG-RAST and similar big data applications demanding elastic computing resources.

Gene sequencing instruments are producing huge volumes of data, straining the capabilities of current database searching algorithms and hindering efforts of researchers analyzing large collections of data to obtain greater insights. In the space of parallel genomic sequence search, most of the popular software packages, like mpiBLAST, use the database segmentation approach, wherein the entire database is sharded and searched on different nodes. However this approach does not scale well with the increasing length of individual query sequences as well as the rapid growth in size of sequence databases. In this paper, we propose a fine-grained parallelism technique, called Orion, that divides the input query into an adaptive number of fragments and shards the database. Our technique achieves higher parallelism (and hence speedup) and load balancing than database sharding alone, while maintaining 100% accuracy. We show that it is 12.3X faster than mpiBLAST for solving a relevant comparative genomics problem.

Virtualization and cloud computing technologies now make it possible to create scalable and reliable virtual high performance computing clusters. Integrating these technologies, however, is complicated by fundamental and inherent differences in the way in which these systems allocate resources to computational tasks. Cloud computing systems immediately allocate available resources or deny requests. In contrast, parallel computing systems route all requests through a queue for future resource allocation. This divergence of allocation policies hinders efforts to implement efficient, responsive, and reliable virtual clusters.