Real-time stereo matching methods primarily focus on enhancing in-domain performance but often overlook the critical importance of generalization in real-world applications. In contrast, recent stereo foundation models leverage monocular foundation models (MFMs) to improve generalization, but typically suffer from substantial inference latency. To address this trade-off, we propose Generalized Geometry Encoding Volume (GGEV), a novel real-time stereo matching network that achieves strong generalization. We first extract depth-aware features that encode domain-invariant structural priors as guidance for cost aggregation. Subsequently, we introduce a Depth-aware Dynamic Cost Aggregation (DDCA) module that adaptively incorporates these priors into each disparity hypothesis, effectively enhancing fragile matching relationships in unseen scenes. Both steps are lightweight and complementary, leading to the construction of a generalized geometry encoding volume with strong generalization capability. Experimental results demonstrate that our GGEV surpasses all existing real-time methods in zero-shot generalization capability, and achieves state-of-the-art performance on the KITTI 2012, KITTI 2015, and ETH3D benchmarks.

Memory caches are being used aggressively in today's data analytics systems such as Spark, Tez, and Piccolo. The significant performance impact of caches and their limited sizes call for efficient cache management in data analytics clusters. However, prevalent data analytics systems employ rather simple cache management policies—notably Least Recently Used (LRU) and Least Frequently Used (LFU)—that areobliviousto the application semantics of data dependency, expressed as directed acyclic graphs (DAGs). Without this knowledge, cache management can, at best, be performed by “guessing” the future data access patterns based on history, which frequently results in inefficient, erroneous caching with a low hit rate and a long response time. Worse still, the lack of data dependency knowledge makes it impossible to retain theall-or-nothingcache property of cluster applications, in that a compute task cannot be sped up unless all the dependent data has been kept in the main memory. In this paper, we propose a novel cache replacement policy, named Least Reference Count (LRC), which exploits the application's data dependency information to optimize the cache management. LRC keeps track of thereference countof each data block, defined as the number of dependent child blocks that have not been computed yet, and always evicts the block with the smallest reference count. Furthermore, we incorporate the all-or-nothing requirement into LRC by coordinately managing the reference counts of all the input data blocks for the same computation. We demonstrate the efficacy of LRC through both empirical analysis and cluster deployments against popular benchmarking workloads. Our Spark implementation shows that, the proposed policies well address the all-or-nothing requirement and significantly improve the cache performance. Compared with LRU and a recently proposed caching policy called MEMTUNE, LRC improves the caching performance of typical workloads in production clusters by 22 and 284 percent, respectively.

The remarkable advances of Machine Learning (ML) have spurred an increasing demand for ML- as-a-Service on public cloud: developers train and publish ML models as online services to provide low-latency inference for dynamic queries. The primary challenge of ML model serving is to meet the response-time Service-Level Objectives (SLOs) of inference workloads while minimizing serving cost. In this article, we proposes MArk (Model Ark), a general-purpose inference serving system, to tackle the dual challenge of SLO compliance and cost effectiveness. MArk employs three design choices tailored to inference workload. First, MArk dynamically batches requests and opportunistically serves them using expensive hardware accelerators (e.g., GPU) for improved performance-cost ratio. Second, instead of relying on feedback control scaling or over-provisioning to serve dynamic workload, which can be too slow or too expensive, MArk employs predictive autoscaling to hide the provisioning latency at low cost. Third, given the stateless nature of inference serving, MArk exploits the flexible, yet costly serverless instances to cover occasional load spikes that are hard to predict. We evaluated the performance of MArk using several state-of-the-art ML models trained in TensorFlow, MXNet, and Keras. Compared with the premier industrial ML serving platform SageMaker, MArk reduces the serving cost up to$7.8\times$while achieving even better latency performance.

Many organizations own data but have limited machine learning expertise (data owners). On the other hand, organizations that have expertise need data from diverse sources to train truly generalizable models (model owners). With the advancement of machine learning (ML) and its growing awareness, the data owners would like to pool their data and collaborate with model owners, such that both entities can benefit from the obtained models. In such a collaboration, the data owners want to protect the privacy of its training data, while the model owners desire the confidentiality of the model and the training method that may contain intellectual properties. Existing private ML solutions, such as federated learning and split learning, cannot simultaneously meet the privacy requirements of both data and model owners.

Machine-Learning-as-a-Service (MLaaS) enables practitioners and AI service providers to train and deploy ML models in the cloud using diverse and scalable compute resources. A common problem for MLaaS users is to choose from a variety of training deployment options, notably scale-up (using more capable instances) and scale-out (using more instances), subject to the budget limits and/or time constraints. State-of-the-art (SOTA) approaches employ analytical modeling for finding the optimal deployment strategy. However, they have limited applicability as they must be tailored to specific ML model architectures, training framework, and hardware. To quickly adapt to the fast evolving design of ML models and hardware infrastructure, we propose a new Bayesian Optimization (BO) based method HeterBO for exploring the optimal deployment of training jobs. Unlike the existing BO approaches for general applications, we consider the heterogeneous exploration cost and machine learning specific prior to significantly improve the search efficiency. This paper culminates in a fully automated MLaaS training Cloud Deployment system (MLCD) driven by the highly efficient HeterBO search method. We have extensively evaluated MLCD in AWS EC2, and the experimental results show that MLCD outperforms two SOTA baselines, conventional BO and CherryPick, by 3.1× and 2.34×, respectively.

Large machine learning models are typically trained in parallel and distributed environments. The model parameters are iteratively refined by multiple worker nodes in parallel, each processing a subset of the training data. In practice, the training is usually conducted in an asynchronous parallel manner, where workers can proceed to the next iteration before receiving the latest model parameters. While this maximizes the rate of updates, the price paid is compromised training quality as the computation is usually performed using stale model parameters. To address this problem, we propose a new scheme, termed speculative synchronization. Our scheme allows workers to speculate about the recent parameter updates from others on the fly, and if necessary, the workers abort the ongoing computation, pull fresher parameters, and start over to improve the quality of training. We design an effective heuristic algorithm to judiciously determine when to restart training iterations with fresher parameters by quantifying the gain and loss. We implement our scheme in MXNet-a popular machine learning framework-and demonstrate its effectiveness through cluster deployment atop Amazon EC2. Experimental results show that speculative synchronization achieves up to 3× speedup over the asynchronous parallel scheme in many machine learning applications, with little additional communication overhead.

Fast track is a software speculation system that enables unsafe optimization of sequential code. It speculatively runs optimized code to improve performance and then checks the correctness of the speculative code by running the original program on multiple processors. We present the interface design and system implementation for Fast Track. It lets a programmer or a profiling tool mark fast-track code regions and uses a run-time system to manage the parallel execution of the speculative process and its checking processes and ensures the correct display of program outputs. The core of the run-time system is a novel concurrent algorithm that balances exploitable parallelism and available processors when the fast track is too slow or too fast. The programming interface closely affects the run-time support. Our system permits both explicit and implicit end markers for speculatively optimized code regions as well as extensions that allow the use of multiple tracks and user designed correctness checking. We discuss the possible uses of speculative optimization and demonstrate the effectiveness of our prototype system by examples of unsafe semantic optimization and a general system for fast memory-safety checking, which is able to reduce the checking time by factors between 2 and 7 for large sequential code on a 8-CPU system.

Good spatial locality alleviates both the latency and bandwidth problem of memory by boosting the effect of prefetching and improving the utilization of cache. However, conventional definitions of spatial locality are inadequate for a programmer to precisely quantify the quality of a program, to identify causes of poor locality, and to estimate the potential by which spatial locality can be improved. This paper describes a new, component-based model for spatial locality. It is based on measuring the change of reuse distances as a function of the data-block size. It divides spatial locality into components at program and behavior levels. While the base model is costly because it requires the tracking of the locality of every memory access, the overhead can be reduced by using small inputs and by extending a sampling-based tool. The paper presents the result of the analysis for a large set of benchmarks, the cost of the analysis, and the experience of a user study, in which the analysis helped to locate a data-layout problem and improve performance by 7 % with a 6-line change in an application with over 2,000 lines.

Software transactional memory systems enable a programmer to easily write concurrent data structures such as lists, trees, hashtables, and graphs, where non-conflicting operations proceed in parallel. Many of these structures take the abstract form of a dictionary, in which each transaction is associated with a search key. By regrouping transactions based on their keys, one may improve locality and reduce conflicts among parallel transactions. In this paper, we present an executor that partitions transactions among available processors. Our key-based adaptive partitioning monitors incoming transactions, estimates the probability distribution of their keys, and adaptively determines the (usually nonuniform) partitions. By comparing the adaptive partitioning with uniform partitioning and round-robin keyless partitioning on a 16-processor SunFire 6800 machine, we demonstrate that key-based adaptive partitioning significantly improves the throughput of finegrained parallel operations on concurrent data structures.

Many sequential applications are difficult to parallelize because of unpredictable control flow, indirect data access, and input-dependent parallelism. These difficulties led us to build a software system for behavior oriented parallelization (BOP), which allows a program to be parallelized based on partial information about program behavior, for example, a user reading just part of the source code, or a profiling tool examining merely one or few executions.

More and more powerful computer technology inspires people to investigate information hidden under huge amounts of documents. In this report, we are especially interested in documents with relative time order, which we also call document streams. Examples include TV news, forums, emails of company projects, call center telephone logs, etc. To get an insight into these document streams, first we need to detect the events among the document streams. We use a time-sensitive Dirichlet process mixture model to find the events in the document streams. A time sensitive Dirichlet process mixture model is a generative model, which allows a potentially infinite number of mixture components and uses a Dirichlet compound multinomial model to model the distribution of words in documents. In this report, we consider three different time sensitive Dirichlet process mixture models: an exponential decay kernel model, a polynomial decay function kernel Dirichlet process model and a sliding window kernel model. Experiments on the TDT2 dataset have shown that the time sensitive models perform 18-20% better in terms of accuracy than the Dirichlet process mixture model. The sliding windows kernel and the polynomial kernel are more promising in detecting events. We use ThemeRiver to provide a visualization of the events along the time axis. With the help of ThemeRiver, people can easily get an overall picture of how different events evolve. Besides ThemeRiver, we investigate using top words as a high-level summarization of each event. Experiment results on TDT2 dataset suggests that the sliding window kernel is a better choice both in terms of capturing the trend of the events and expressibility

In classic pattern recognition problems, classes are mutually exclusive by definition. However, in many applications, it is quite natural that some instances belong to multiple classes at the same time. In other words, these applications are multi-labeled, classes are overlapped by definition and each instance may be associated to multiple classes. In this paper, we present a comparative study on various multi-label approaches using both gene and scene data sets. We expect our research efforts provide useful insights on the relationships among various classifiers as well as various evaluation measures and shed lights on future research. Although there is no clear winner across various performance measures, SVM binary and multi-label ADTree perform better than the others on most counts. We then propose a meta-learning approach by combining SVM binary and ADTree. Our experiments demonstrate that the combined method can take the advantages of the single approaches

Most application's performance is impacted by the amount of available memory. In a traditional application, which has a fixed working set size, increasing memory has a beneficial effect up until the application's working set is met. In the presence of garbage collection this relationship becomes more complex. While increasing the size of the program's heap reduces the frequency of collections, collecting a heap with memory paged to the backing store is very expensive. We first demonstrate the presence of an optimal heap size for a number of applications running on a machine with a specific configuration. We then introduce a scheme which adaptively finds this good heap size. In this scheme, we track the memory usage and number of page faults at a program's phase boundaries. Using this information, the system selects the soft heap size. By adapting itself dynamically, our scheme is independent of the underlying main memory size, code optimizations, and garbage collection algorithm. We present several experiments on real applications to show the effectiveness of our approach. Our results show that program-level heap control provides up to a factor of 7.8 overall speedup versus using the best possible fixed heap size controlled by the virtual machine on identical garbage collectors.

In POPL 2002, Petrank and Rawitz showed a universal result— finding optimal data placement is not only NP-hard but also impossible to approximate within a constant factor if P ̸ = NP. Here we study a recently published concept called reference affinity, which characterizes a group of data that are always accessed together in computation. On the theoretical side, we give the complexity for finding reference affinity in program traces, using a novel reduction that converts the notion of distance into satisfiability. We also prove that reference affinity automatically captures the hierarchical locality in divide-and-conquer computations including matrix solvers and N-body simulation. The proof establishes formal links between computation patterns in time and locality relations in space. On the practical side, we show that efficient heuristics exist. In particular, we present a sampling method and show that it is more effective than the previously published technique, especially for data that are often but not always accessed together. We show the effect on generated and real traces. These theoretical and empirical results demonstrate that effective data placement is still attainable in general-purpose programs because common (albeit not all) locality patterns can be precisely modeled and efficiently analyzed.

This paper studies the problem of building multiclass classifiers for tissue classification based on gene expression. The recent development of microarray technologies has enabled biologists to quantify gene expression of tens of thousands of genes in a single experiment. Biologists have begun collecting gene expression for a large number of samples. One of the urgent issues in the use of microarray data is to develop methods for characterizing samples based on their gene expression. The most basic step in the research direction is binary sample classification, which has been studied extensively over the past few years. This paper investigates the next step-multiclass classification of samples based on gene expression. The characteristics of expression data (e.g. large number of genes with small sample size) makes the classification problem more challenging. The process of building multiclass classifiers is divided into two components: (i) selection of the features (i.e. genes) to be used for training and testing and (ii) selection of the classification method. This paper compares various feature selection methods as well as various state-of-the-art classification methods on various multiclass gene expression datasets. Our study indicates that multiclass classification problem is much more difficult than the binary one for the gene expression datasets. The difficulty lies in the fact that the data are of high dimensionality and that the sample size is small. The classification accuracy appears to degrade very rapidly as the number of classes increases. In particular, the accuracy was very low regardless of the choices of the methods for large-class datasets (e.g. NCI60 and GCM). While increasing the number of samples is a plausible solution to the problem of accuracy degradation, it is important to develop algorithms that are able to analyze effectively multiple-class expression data for these special datasets.