Deep learning now offers state-of-the-art accuracy for many prediction tasks. A form of deep learning called deep convolutional neural networks (CNNs) are especially popular on image, video, and time series data. Due ...
详细信息
Deep learning now offers state-of-the-art accuracy for many prediction tasks. A form of deep learning called deep convolutional neural networks (CNNs) are especially popular on image, video, and time series data. Due to its high computational cost, CNN inference is often a bottleneck in analytics tasks on such data. Thus, a lot of work in the computer architecture, systems, and compilers communities study how to make CNN inference faster. In this work, we show that by elevating the abstraction level and re-imagining CNN inference as queries, we can bring to bear database-style queryoptimization techniques to improve CNN inference efficiency. We focus on tasks that perform CNN inference repeatedly on inputs that are only slightly different. We identify two popular CNN tasks with this behavior: occlusion-based explanations (OBE) and object recognition In videos (ORV). OBE is a popular method for "explaining" CNN predictions. It outputs a heatmap over the input to show which regions (e.g., image pixels) mattered most for a given prediction. It leads to many re-inference requests on locally modified inputs. ORV uses CNNs to identify and track objects across video frames. It also leads to many re-inference requests. We cast such tasks in a unified manner as a novel instance of the incremental view maintenance problem and create a comprehensive algebraic framework for incremental CNN inference that reduces computational costs. We produce materialized views of features produced inside a CNN and connect them with a novel multi-query optimization scheme for CNN re-inference. Finally, we also devise novel OBE-specific and ORV-specific approximate inference optimizations exploiting their semantics. We prototype our ideas in Python to create a tool called KRYPTON that supports both CPUs and CPUs. Experiments with real data and CNNs show that KRYPTON reduces runtimes by up to 5x (respectively, 35x) to produce exact (respectively, high-quality approximate) results without raisin
Data integration has evolved to provide efficient data management across distributed and heterogeneous data sources in grid ***, existing works in data integration consider little knowledge about the above application...
详细信息
Data integration has evolved to provide efficient data management across distributed and heterogeneous data sources in grid ***, existing works in data integration consider little knowledge about the above applications. In such settings, queries from the same application are processed independently.
In this paper, application properties are noticed in order to improve query performance for data integration. We present a general-purpose and wellmodular architecture for addressing data integration in grid (DIG) environment first, the modules of which can be flexibly deployed to adapt specific ***, special attention has been paid on query processing to accommodate application workflow with the component of query Flow Processor. It introduces multi-query optimization techniques to speed up overall response time for DIG.
In this paper, we address one of the wireless sensor network query processing issues posed due to the lack of support for multiple sensor network queries. The objective of the paper is to provide efficient and effecti...
详细信息
In this paper, we address one of the wireless sensor network query processing issues posed due to the lack of support for multiple sensor network queries. The objective of the paper is to provide efficient and effective support to multiple queries so that the set of queries are pre-processed before disseminating them into the sensor network. It is very important that only necessary works will be assigned to the sensor network by virtue of strict energy constraint. The problem is modeled by Minimum Set Cover, which is one of the NP-complete problems. We propose an optimization scheme called TAMPA - a Tabu search-based multiple queries optimization to find an optimal merge order. The final set of queries to be sent into the network then can be derived from that merge order. We evaluate the proposed algorithm by conducting extensive simulation studies. The results show that energy can be significantly saved while the overall workload still satisfies the user requirements.
Recent research efforts in the fields of data stream processing show the increasing importance of processing data streams, e.g., in the e-science domain. Together with the advent of peer-to-peer (P2P) networks and gri...
详细信息
Recent research efforts in the fields of data stream processing show the increasing importance of processing data streams, e.g., in the e-science domain. Together with the advent of peer-to-peer (P2P) networks and grid computing, this leads to the necessity of developing new techniques for distributing and processing continuous queries over data streams in such networks. These systems often have to process multiple similar but different continuous aggregation queries simultaneously. Since executing each query separately can lead to significant scalability and performance problems, it is vital to share resources by exploiting similarities in the queries. The challenge is to identify overlapping computations that may not be obvious in the queries themselves. In this paper, we propose a novel algorithmic solution for problem of finding the minimum number of queries in such a distributed-streams setting, in order to optimize the communicate cost across the network. The experiment result show that our approach gives us as much as magnitude performance improvement over the no-share settings.
暂无评论