Work

Data-driven Performance Optimization for Data-intensive Applications

Year: 2024
Type: dissertation
Abstract: Data-intensive applications have attracted considerable attention from researchersin information sciences and enterprises, as these applications have made\nevolutionary breakthroughs in scientific fields and are extremely valuable to produce\nproductivity in businesses. Recently, as the high speed growth of the new generated\ndata, researchers have begun to leverage the useful knowledge hidden in such\nhuge volume of data to optimize the performance of the data-intensive applications.\nHowever, optimize the performance of the data-intensive applications based on the\ndata-driven approaches are still need to be explored.\nIn this thesis, we focus on data-driven performance optimization for dataintensive\napplications. We first study an application, auto-labeling data on mobile\ndevices. How to accurately and efficiently label data on a mobile device is critical\nfor the success of training machine learning models on mobile devices. Auto-labeling\ndata for data-intensive applications on mobile devices is a challenging task, because\ndata is incrementally generated and there is a possibility of having unknown labels\namong new coming data. Furthermore, the rich hardware heterogeneity on mobile\ndevices creates challenges on efficiently executing the autolabeling workload. We\nintroduce Flame, an auto-labeling system that can label dynamically generated data\nwith unknown labels. Flame includes an execution engine that efficiently schedules\nand executes auto-labeling workloads on heterogeneous mobile processors. Evaluating\nFlame with six datasets on two mobile devices, we demonstrate that the labeling\naccuracy of Flame is 11.8%, 16.1%, 18.5%, and 25.2% higher than a state-of-the-art\nlabeling method, transfer learning, semi-supervised learning, and boosting methods\nrespectively. Flame is also energy efficient, it consumes only 328.65mJ and 414.84mJ\nwhen labeling 500 data instances on Samsung S9 and Google Pixel2 respectively.\nFurthermore, running Flame on mobile devices only brings about 0.75 ms additional\nframe latency which is imperceivable by the users.\nSecond, we explore another data-intensive application, the cardinality estimation\nin database systems. Cardinality estimation is a fundamental and critical\nproblem in databases. Recently, many estimators based on deep learning have been\nproposed to solve this problem and they have achieved promising results. However,\nthese estimators struggle to provide accurate results for complex queries, due to not\ncapturing real inter-column and inter-table correlations. Furthermore, none of these\nestimators contain the uncertainty information about their estimations. In this paper,\nwe present a join cardinality estimator called . learns the correlations across\nall columns and all tables in the database. It also contains the uncertainty information\nof each estimation. Among all studied learned estimators, our results are\npromising: (1) has the smallest model size; (2) It has the fastest inference speed; (3)\nCompared with the state of the art estimator, has 10× faster inference speed, and\nprovides 1.3× ∼ 6.7× smaller estimation errors for complex queries; (4) To the best\nof our knowledge, is the first estimator that incorporates uncertainty information for\ncardinality estimation into a deep learning model.\nFurthermore, we also study the data loading problem for large-scale distributed\ntraining. The resource-hungry and time-consuming process of training Deep\nNeural Networks (DNNs) can be accelerated by optimizing and/or scaling computations\non accelerators such as GPUs. However, the loading and pre-processing of\ntraining samples then often emerges as a new bottleneck. This data loading process\nengages a complex pipeline that extends from the sampling of training data on\nexternal storage to delivery of those data to GPUs, and that comprises not only\nexpensive I/O operations but also decoding, shuffling, batching, augmentation, and\nother operations. We propose in this paper a new holistic approach to data loading\nthat addresses three challenges not sufficiently addressed by other methods: I/O load\nimbalances among the GPUs on a node; rigid resource allocations to data loading and\ndata preprocessing steps, which lead to idle resources and bottlenecks; and limited efficiency\nof caching strategies based on pre-fetching due to eviction of training samples\nneeded soon at the expense of those needed later. We first present a study of key bottlenecks observed as training samples flow through the data loading and preprocessing\npipeline. Then, we describe Lobster, a data loading runtime that uses performance\nmodeling and advanced heuristics to combine flexible thread management with optimized\neviction for distributed caching in order to mitigate I/O overheads and load\nimbalances. Experiments with a range of models and datasets show that the Lobster\napproach reduces both I/O overheads and end-to-end training times by up to 1.5×\ncompared with state-of-the-art approaches.\nFinally, we study the cardinality estimation for string predicates in database\nsystems. Cardinality estimation for string predicates is a notoriously challenging\nproblem in database systems. This paper presents ArbiLIKE, an advanced deep\nlearning-based cardinality estimator for arbitrary LIKE predicates. ArbiLIKE utilizes\na cardinality-aware embedding technique to encode LIKE predicates into feature\nvectors. It further incorporates an innovative sequence model to capture the semantic\ninformation of different substrings, enhancing the estimation accuracy. ArbiLIKE is\nalso capable of handling LIKE predicates with any combination of wildcards (“%”,\n“ ”). Empirical evaluations showcase ArbiLIKE’s promising accuracy, achieving estimation\nerrors that are up to 165.1× smaller than those of eight baselines, including\nstate-of-the-art methods. As a generic estimator, ArbiLIKE realizes error reductions\nranging from 1.4 to 93.1× for LIKE predicates with multiple wildcards in comparison\nto the existing techniques. To the best of our knowledge, ArbiLIKE is the first deep\nlearning-based estimator capable of handling arbitrary LIKE predicates.
Author: Liu, Jie
Language: English
Open Access status: green
OpenAlex ID: W7027634192
Cites: 182
Cited by:
Related to: 10