Abstract Although more layers and more parameters generally improve the accuracy of the models, such big models generally have high computational complexity and require big memory, which exceed the capacity of small devices for inference and incurs long training time. In addition, it is difficult to afford long training time and inference time of big models even in high performance servers, as well. As an efficient approach to compress a large deep model (a teacher model) to a compact model (a student model), knowledge distillation emerges as a promising approach to deal with the big models. Existing knowledge distillation methods cannot exploit the elastic available computing resources and correspond to low efficiency. In this paper, we propose an Elastic Deep Learning framework for knowledge Distillation, that is, EDL‐Dist. The advantages of EDL‐Dist are threefold. First, the inference and the training process is separated. Second, elastic available computing resources can be utilized to improve the efficiency. Third, fault‐tolerance of the training and inference processes is supported. We take extensive experimentation to show that the throughput of EDL‐Dist is up to 3.125 times faster than the baseline method (online knowledge distillation) while the accuracy is similar or higher.

With the advancement and dominant service of Internet videos, the content-based video deduplication system becomes an essential and dependent infrastructure for Internet video service. However, the explosively growing video data on the Internet challenges the system design and implementation for its scalability in several ways. (1) Although the quantization-based indexing techniques are effective for searching visual features at a large scale, the costly re-training over the complete dataset must be done periodically. (2) The high-dimensional vectors for visual features demand increasingly large SSD space, degrading I/O performance. (3) Videos crawled from the Internet are diverse, and visually similar videos are not necessarily the duplicates, increasing deduplication complexity. (4) Most videos are edited ones. The duplicate contents are more likely discovered as clips inside the videos, demanding processing techniques with close attention to details.

Visual contents, including images and videos, are dominant on the Internet today. The conventional search engine is mainly designed for textual documents, which must be extended to process and manage increasingly high volumes of visual data objects. In this paper, we present Mixer, an effective system to identify and analyze visual contents and to extract their features for data retrievals, aiming at addressing two critical issues: (1) efficiently and timely understanding visual contents, (2) retrieving them at high precision and recall rates without impairing the performance. In Mixer, the visual objects are categorized into different classes, each of which has representative visual features. Subsystems for model production and model execution are developed. Two retrieval layers are designed and implemented for images and videos, respectively. In this way, we are able to perform aggregation retrievals of the two types in efficient ways. The experiments with Baidu's production workloads and systems show that Mixer halves the model production time and raises the feature production throughput by 9.14x. Mixer also achieves the precision and recall of video retrievals at 95% and 97%, respectively. Mixer has been in its daily operations, which makes the search engine highly scalable for visual contents at a low cost. Having observed productivity improvement of upper-level applications in the search engine, we believe our system framework would generally benefit other data processing applications.

The freshness of web page indices is the key to improving searching quality of search engines. In Baidu, the major search engine in China, we have developed DirectLoad, an index updating system for efficiently delivering the webscale indices to nationwide data centers. However, the web-scale index updating suffers from increasingly high data volumes during network transmission and inefficient I/O transactions due to slow disk operations. DirectLoad accelerates the index updating streams from two aspects: 1) DirectLoad effectively cuts down the overwhelmingly high volume of indices in transmission by removing the redundant data across versions, and mutates regular operations in a key-value storage system for successful accesses to the deduplicated datasets. 2) DirectLoad significantly improves the I/O efficiency by replacing the LSMTree with a memory-resident table (memtable) and appendingonly-files (AOFs) on disk. Specifically, the write amplification stemming from sorting operations on disk is eliminated, and a lazy garbage collection policy further improves the I/O performance at the software level. In addition, DirectLoad directly manipulates the SSD native interfaces to remove the write amplification at the hardware level. In practice, 63% updating bandwidth has been saved due to the deduplication, and the write throughput to SSDs is increased by 3x. The index updating cycle of our production workloads has been compressed from 15 days to 3 days after deploying DirectLoad. In this paper, we show the effectiveness and efficiency of an in-memory index updating system, which is disruptive to the framework in a conventional memory hierarchy. We hope that this work contributes a strong case study in the system research literature.

Fast data analytics at an increasingly large scale has become a critical task in any Internet service company. For example, in Baidu, the major search engine company in China, large volumes of Web and business data in PB-scale are timely and constantly acquired and analyzed for the purposes of evaluating product revenue, tracking product demanding activities on market, predicting user behavior, upgrading product rankings, and diagnosing spam cases, and many others. Response time for queries of various data analytics not only affects user experiences, but also has a serious impact on productivity of business operations. In this paper, to meet the challenge of fast data analytics, we present Feisu (meaning fast in Chinese), a data integration system over heterogeneous storage systems, which has been widely used in Baidu's critical and daily business analytics applications after our R&D efforts. Feisu is designed and implemented to co-work together with several heterogeneous storage systems, and exploit the query similarity embedded in complex query workloads. Our experiments using real world workloads show that Feisu can significantly improve query performance in Baidu. Feisu has been in production use in Baidu for two years to effectively manage over dozens of petabytes of data for various applications.

We present Fatman, an enterprise-scale archival storage based on volunteer contribution resources from underutilized web servers, usually deployed on thousands of nodes with spare storage capacity. Fatman is specifically designed for enhancing the utilization of existing storage resources and cutting down the hardware purchase cost. Two major concerned issues of the system design are maximizing the resource utilization of volunteer nodes without violating Service Level Objectives (SLOs) and minimizing the cost without reducing the availability of archival system. Fatman has been widely deployed on tens of thousands of server nodes across several datacenters, provided more than 100PB storage capacity and served dozens of internal mass-data applications. The system realizes an efficient storage quota consolidation by strong isolation and budget limitation, to maximally support resources contribution without any degradation on host-level SLOs. It firstly improves data reliability by applying disk failure prediction to minish failure recovery cost, named fault-aware data management, dramatically reduces the MTTR by 76.3% and decreases file crash ratio by 35% on real-life product workload.

In computational grids, a virtual organization (VO) is a dynamic coupling of multiple Linux/Unix nodes for resource sharing under specific polices. Currently, VO support functionalities are generally implemented as grid middleware. However, the usability of grids is often impaired by the complexity of configuring and maintaining a new layer of security infrastructure as well as adapting to new interfaces of security enabled services. In this paper, we present an OS-level approach to provide native VO support functionalities, which is a part of XtreemOS project [18]. Our approach adopts pluggable frameworks existing in current OS as extension points to implement VO support, avoiding modification of kernel codes and easily turning traditional OSes into grid-aware ones. The performance evaluation of NAS parallel benchmarks (NPB) shows that our current implementation incurs trivial overhead on original systems.