Lixin Gao

DPillar: Dual-port server interconnection network for large scale data centers

To meet the huge demands of computation power and storage space, a future data center may have to include up to millions of servers. The conventional hierarchical tree-based data center network architecture faces several challenges in scaling a data center to that size. Previous research effort has shown that a server-centric architecture, where servers are not only computation and storage workstations but also intermediate nodes relaying traffic for other servers, performs well in scaling a data center to a huge number of servers. This paper presents a server-centric data center network called DPillar, whose topology is inspired by the classic butterfly network. DPillar provides several nice properties and achieves the balance between topological scalability, network performance, and cost efficiency, which make it suitable for building large scale future data centers. Using only commodity hardware, a DPillar network can easily accommodate millions of servers. The structure of a DPillar network is symmetric so that any network bottleneck is eliminated at the architectural level. With each server having only two ports, DPillar is able to provide the bandwidth to support communication intensive distributed applications. This paper studies the interconnection features of DPillar, how to compute routes in DPillar, and how to forward packets in DPillar. Extensive simulation experiments have been performed to evaluate the performance of DPillar. The results show that DPillar performs well even in the presence of a large number of server and switch failures.

Scalable Nonnegative Matrix Factorization with Block-wise updates

Nonnegative Matrix Factorization (NMF) has been applied with great success to many applications. As NMF is applied to massive datasets such as web-scale dyadic data, it is desirable to leverage a cluster of machines to speed up the factorization. However, it is challenging to efficiently implement NMF in a distributed environment. In this paper, we show that by leveraging a new form of update functions, we can perform local aggregation and fully explore parallelism. Moreover, under the new form of update functions, we can perform frequent updates, which aim to use the most recently updated data whenever possible. As a result, frequent updates are more efficient than their traditional concurrent counterparts. Through a series of experiments on a local cluster as well as the Amazon EC2 cloud, we demonstrate that our implementation with frequent updates is up to two orders of magnitude faster than the existing implementation with the traditional form of update functions.

Accelerating Expectation-Maximization Algorithms with Frequent Updates

Expectation Maximization is a popular approach for parameter estimation in many applications such as image understanding, document classification, or genome data analysis. Despite the popularity of EM algorithms, it is challenging to efficiently implement these algorithms in a distributed environment. In particular, many EM algorithms that frequently update the parameters have been shown to be much more efficient than their concurrent counterparts. Accordingly, we propose two approaches to parallelize such EM algorithms in a distributed environment so as to scale to massive data sets. We prove that both approaches maintain the convergence properties of the EM algorithms. Based on the approaches, we design and implement a distributed framework, FreEM, to support the implementation of frequent updates for the EM algorithms. We show its efficiency through three well-known EM applications: k-means clustering, fuzzy c-means clustering and parameter estimation for the Gaussian Mixture model. We evaluate our framework on both a local cluster of machines and the Amazon EC2 cloud. Our evaluation shows that the EM algorithms with frequent updates implemented on FreEM can run much faster than those implementations with traditional concurrent updates.

Scalable Distributed Belief Propagation with Prioritized Block Updates

Belief propagation (BP) is a popular method for performing approximate inference on probabilistic graphical models. However, its message updates are time-consuming, and the schedule for updating messages is crucial to its running time and even convergence. In this paper, we propose a new scheduling scheme that selects a set of messages to update at a time and leverages a novel priority to determine which messages are selected. Additionally, an incremental update approach is introduced to accelerate the computation of the priority. As the size of the model grows, it is desirable to leverage the parallelism of a cluster of machines to reduce the inference time. Therefore, we design a distributed framework, Prom, to facilitate the implementation of BP algorithms. We evaluate the proposed scheduling scheme (supported by Prom) via extensive experiments on a local cluster as well as the Amazon EC2 cloud. The evaluation results show that our scheduling scheme outperforms the state-of-the-art counterpart.

Efficient analytics on ordered datasets using MapReduce

Efficiently analyzing data on a large scale can be vital for data owners to gain useful business intelligence. One of the most common datasets used to gain business intelligence is event log files. Oftentimes, records in event log files that are time sorted, need to be grouped by user ID or transaction ID in order to mine user behaviors, such as click through rate, while preserving the time order. This kind of analytical workload is here referred to as RElative Order-pReserving based Grouping (Re-Org). Using MapReduce/Hadoop, a popular big data analysis tool, in an as-is manner for executing Re-Org tasks on ordered datasets is not efficient due to its internal sort-merge mechanism. We propose a framework that adopts an efficient group-order-merge mechanism to provide faster execution of Re-Org tasks and implement it by extending Hadoop. Experimental results show a 2.2x speedup over executing Re-Org tasks in plain vanilla Hadoop.

Efficient social network data query processing on MapReduce

Social network data analysis becomes increasingly important for business intelligence and online social services. Lots of social network data is presented by Resource Description Framework (RDF). Accordingly, SPARQL, an RDF query language, becomes popular for social network data analysis. As the sizes of social networks expand rapidly, a SPARQL query usually involves a large quantity of data, and thus parallelizing its execution is desirable. MapReduce is a well-known and popular big data analysis tool. However, the state-of-the-art translation from SPARQL queries to MapReduce jobs is not efficient because it mainly follows a two layer rule which needs to transform the SPARQL triple pattern to the standard SQL join. In this paper, we propose two primitives to enable efficient translation from SPARQL queries to MapReduce jobs. We use multiple-join-with-filter to substitute traditional SQL multiple join when feasible, and merge different stages in the query workflow. The evaluation on social network data benchmarks shows that the translation based on these two primitives can achieve up to 2x speedup in query running time comparing to the traditional two layer scheme.

Real-time volume control for interactive network traffic replay

Traffic volume control is one of the fundamental requirements in traffic generation and transformation. However, due to the complex interactions between the generated traffic and replay environment (delay, packet loss, connection blocking, etc), controlling traffic volume in interactive network traffic replay becomes a challenging problem. In this paper, we present a novel model-based analytical method to address this problem where the generated traffic volume is regulated through adjustment of input traffic volume. By analyzing the replay mechanism in terms of how packets are processed, and properly choosing buffered packets amount and to-be-received packets amount as system states, we present a novel model-based analytical method to obtain the desired input volume. The traffic volume control problem is then converted to a state prediction problem where we employ Recursive Least Square (RLS) filter to predict system states. As compared to other adaptive control techniques, our method does not involve any learning scheme and hence completely requires no convergence time. Experimental studies further indicate that our method is efficient in tracking target traffic volume (both static and time-varying) and works under a wide range of network conditions.

Co-ClusterD: A Distributed Framework for Data Co-Clustering with Sequential Updates

Co-clustering has emerged to be a powerful data mining tool for two-dimensional co-occurrence and dyadic data. However, co-clustering algorithms often require significant computational resources and have been dismissed as impractical for large data sets. Existing studies have provided strong empirical evidence that expectation-maximization (EM) algorithms (e.g., k-means algorithm) with sequential updates can significantly reduce the computational cost without degrading the resulting solution. Motivated by this observation, we introduce sequential updates for alternate minimization co-clustering (AMCC) algorithms which are variants of EM algorithms, and also show that AMCC algorithms with sequential updates converge. We then propose two approaches to parallelize AMCC algorithms with sequential updates in a distributed environment. Both approaches are proved to maintain the convergence properties of AMCC algorithms. Based on these two approaches, we present a new distributed framework, Co-ClusterD, which supports efficient implementations of AMCC algorithms with sequential updates. We design and implement Co-ClusterD, and show its efficiency through two AMCC algorithms: fast nonnegative matrix tri-factorization (FNMTF) and information theoretic co-clustering (ITCC). We evaluate our framework on both a local cluster of machines and the Amazon EC2 cloud. Empirical results show that AMCC algorithms implemented in Co-ClusterD can achieve a much faster convergence and often obtain better results than their traditional concurrent counterparts.

Asynchronous Distributed Incremental Computation on Evolving Graphs

Graph algorithms have become an essential component in many real-world applications. An essential property of graphs is that they are often dynamic. Many applications must update the computation result periodically on the new graph so as to keep it up-to-date. Incremental computation is a promising technique for this purpose. Traditionally, incremental computation is typically performed synchronously, since it is easy to implement. In this paper, we illustrate that incremental computation can be performed asynchronously as well. Asynchronous incremental computation can bypass synchronization barriers and always utilize the most recent values, and thus it is more efficient than its synchronous counterpart. Furthermore, we develop a distributed framework, GraphIn, to facilitate implementations of incremental computation on massive evolving graphs. We evaluate our asynchronous incremental computation approach via extensive experiments on a local cluster as well as the Amazon EC2 cloud. The evaluation results show that it can accelerate the convergence speed by as much as 14x when compared to recomputation from scratch.

A Scalable Distributed Framework for Efficient Analytics on Ordered Datasets

One of the most common datasets used by many corporations to gain business intelligence is event log files. Oftentimes, the records in event log files are temporally ordered, and need to be grouped by user ID with the temporal ordering preserved to facilitate mining user behaviors. This kind of analytical workload, here referred to as Relative Order-preserving based Grouping (RE-ORG), is quite common in big data analytics. Using MapReduce/Hadoop for executing RE-ORG tasks on ordered datasets is not efficient due to its internal sort-merge mechanism. In this paper, we propose a distributed framework that adopts an efficient group-order-merge mechanism to provide faster execution of RE-ORG tasks. We demonstrate the advantage of our framework by comparing its performance with Hadoop through extensive experiments on real-world datasets. The evaluation results show that our framework can achieve up to 6.3× speedup over Hadoop in executing RE-ORG tasks.