David R. Swanson

An efficient fault-tolerant scheduling algorithm for real-time tasks with precedence constraints in heterogeneous systems

In this paper, we investigate an efficient off-line scheduling algorithm in which real-time tasks with precedence constraints are executed in a heterogeneous environment. It provides more features and capabilities than existing algorithms that schedule only independent tasks in real-time homogeneous systems. In addition, the proposed algorithm takes the heterogeneities of computation, communication and reliability into account, thereby improving the reliability. To provide fault-tolerant capability, the algorithm employs a primary-backup copy scheme that enables the system to tolerate permanent failures in any single processor. In this scheme, a backup copy is allowed to overlap with other backup copies on the same processor, as long as their corresponding primary copies are allocated to different processors. Tasks are judiciously allocated to processors so as to reduce the schedule length as well as the reliability cost, defined to be the product of processor failure rate and task execution time. In addition, the time for detecting and handling a permanent fault is incorporated into the scheduling scheme, thus making the algorithm more practical. To quantify the combined performance of fault-tolerance and schedulability, the performability measure is introduced Compared with the existing scheduling algorithms in the literature, our scheduling algorithm achieves an average of 16.4% improvement in reliability and an average of 49.3% improvement in performability.

Matchmaking: A New MapReduce Scheduling Technique

MapReduce is a powerful platform for large-scale data processing. To achieve good performance, a MapReduce scheduler must avoid unnecessary data transmission by enhancing the data locality (placing tasks on nodes that contain their input data). This paper develops a new MapReduce scheduling technique to enhance map tasks data locality. We have integrated this technique into Hadoop default FIFO scheduler and Hadoop fair scheduler. To evaluate our technique, we compare not only MapReduce scheduling algorithms with and without our technique but also with an existing data locality enhancement technique (i.e., the delay algorithm developed by Face book). Experimental results show that our technique often leads to the highest data locality rate and the lowest response time for map tasks. Furthermore, unlike the delay algorithm, it does not require an intricate parameter tuning process.

Continuum predictions from molecular dynamics simulations: Shock waves

Techniques are investigated for obtaining continuously distributed local properties from the positions and velocities of constituent atoms. A localization function is used to calculate the local density, temperature, and velocity from the results of molecular dynamics simulations of shock waves in a two-dimensional model system. The two-dimensional spatial variations of the local properties are found, and the width of the localization function is varied to optimize the presentation. A vector plot of the local velocity shows turbulence behind the shock with a resolution on the order of a nanometer. Contour plots show the mass density and local temperature at a hot spot caused by a collapsed void.

An agent-based infrastructure for parallel Java on heterogeneous clusters

In this paper, we introduce an agent-based infrastructure that provides software services and functions for developing and deploying high performance programming models and applications on clusters. A Java-based prototype, based on this architecture, has been developed. Since this system is written completely in Java, it is portable and allows executing programs in parallel across multiple heterogeneous platforms. With the agent-based infrastructure, users need not deal with the mechanisms of deploying and loading user classes on the heterogeneous cluster. Moreover, details of scheduling, controlling, monitoring, and executing user jobs are hidden. In addition, the management of system resources is made transparent to the user. Such uniform services, when rendered available in a ubiquitous manner, are essential for facilitating the development and deployment of scalable high performance Java applications on clusters. An initial deployment over a heterogeneous, distributed cluster results in significantly enhanced performance; absolute performance compared to C (MPI) improves with increased granularity of the algorithms.

Modeling parallel applications performance on heterogeneous systems

The current technologies have made it possible to execute parallel applications across heterogeneous platforms. However, the performance models available do not provide adequate methods to calculate, compare and predict the applications performance on these platforms. In this paper, we discuss an enhanced performance evaluation model for parallel applications on heterogeneous systems. In our analysis, we include machines of different architectures, specifications and operating environments. We also discuss the enabling technologies that facilitate such heterogeneous applications. The model is then validated through experimental measurements using an agent-based parallel Java system, which facilitates simultaneous utilization of heterogeneous systems for parallel applications. The model provides good evaluation metrics that allow developers to assess and compare the parallel heterogeneous applications performances.

Dynamic Load Balancing for I/O-Intensive Tasks on Heterogeneous Clusters

Since I/O-intensive tasks running on a heterogeneous cluster need a highly effective usage of global I/O resources, previous CPU- or memory-centric load balancing schemes suffer significant performance drop under I/O-intensive workload due to the imbalance of I/O load. To solve this problem, we develop two I/O-aware load-balancing schemes, which consider system heterogeneity and migrate more I/O-intensive tasks from a node with high I/O utilization to those with low I/O utilization. If the workload is memory-intensive in nature, the new method applies a memory-based load balancing policy to assign the tasks. Likewise, when the workload becomes CPU-intensive, our scheme leverages a CPU-based policy as an efficient means to balance the system load. In doing so, the proposed approach maintains the same level of performance as the existing schemes when I/O load is low or well balanced. Results from a trace-driven simulation study show that, when a workload is I/O-intensive, the proposed schemes improve the performance with respect to mean slowdown over the existing schemes by up to a factor of 8. In addition, the slowdowns of almost all the policies increase consistently with the system heterogeneity.

A dynamic load balancing scheme for I/O-intensive applications in distributed systems

In this paper, a new I/O-aware load-balancing scheme is presented to improve overall performance of a distributed system with a general and practical workload including I/O activities. The proposed scheme dynamically detects I/O load imbalance on nodes of a distributed system and determines whether to migrate the I/O requests of some jobs from overloaded nodes to other less- or under-loaded nodes, depending on data migration cost and remote I/O access overhead. Besides balancing I/O load, the scheme judiciously takes into account both CPU and memory load sharing in distributed systems, thereby maintaining the same level of performance as the existing schemes when I/O load is low or well balanced. Results from a trace-driven simulation study show that, compared with the existing schemes that only consider CPU and memory, the proposed scheme reduces the mean slowdown by up to 54.5% (with an average of 39.9%). On the other hand, when compared to the existing approaches that only consider I/O, the proposed scheme reduces the mean slowdown by up to 57.2% (with an average of 31.6%). More importantly, the new scheme improves over a very recent algorithm found in the literature that considers all the three resources by up to 49.6% (with an average of up to 41.9%).

Design, implementation and performance evaluation of a cost-effective, fault-tolerant parallel virtual file system

Fault tolerance is one of the most important issues for parallel file systems. This paper presents the design, implementation and performance evaluation of a cost-effective, fault-tolerant parallel virtual file system (CEFT-PVFS) that provides parallel I/O service without requiring any additional hardware by utilizing existing commodity disks on cluster nodes and incorporates fault tolerance in the form of disk mirroring. While mirroring is a straightforward idea, we have implemented this open source system and conducted extensive experiments to evaluate the feasibility, efficiency and scalability of this fault tolerant approach on one of the current largest clusters, where the issues of data consistency and recovery are also investigated. Four mirroring protocols are proposed, reflecting whether the fault-tolerant operations are client driven or server driven; synchronous or asynchronous. Their relative merits are assessed by comparing their write performances, measured in the real systems, and their reliability and availability measures, obtained through analytical modeling. The results indicate that, in cluster environments, mirroring can improve the reliability by a factor of over 40 (4000%) while sacrificing the peak write performance by 33--58% when both systems are of identical sizes (i.e., counting the 50% mirroring disks in the mirrored system). In addition, protocols with higher peak write performance are less reliable than those with lower peak write performance, with the latter achieving a higher reliability and availability at the expense of some write bandwidth. A hybrid protocol is proposed to optimize this tradeoff.

HOG: Distributed Hadoop MapReduce on the Grid

MapReduce is a powerful data processing platform for commercial and academic applications. In this paper, we build a novel Hadoop MapReduce framework executed on the Open Science Grid which spans multiple institutions across the United States - Hadoop On the Grid (HOG). It is different from previous MapReduce platforms that run on dedicated environments like clusters or clouds. HOG provides a free, elastic, and dynamic MapReduce environment on the opportunistic resources of the grid. In HOG, we improve Hadoops fault tolerance for wide area data analysis by mapping data centers across the U.S. to virtual racks and creating multi-institution failure domains. Our modifications to the Hadoop framework are transparent to existing Hadoop MapReduce applications. In the evaluation, we successfully extend HOG to 1100 nodes on the grid. Additionally, we evaluate HOG with a simulated Facebook Hadoop MapReduce workload. We conclude that HOGs rapid scalability can provide comparable performance to a dedicated Hadoop cluster.

JOPI: a Java object‐passing interface

Recently there has been an increasing interest in developing parallel programming capabilities in Java to harness the vast resources available in clusters, grids and heterogeneous networked systems. In this paper, we introduce a Java object‐passing interface (JOPI) library. JOPI provides Java programmers with the necessary functionality to write object‐passing parallel programs in distributed heterogeneous systems. JOPI provides a Message Passing Interface (MPI)‐like interface that can be used to exchange objects among processes. In addition to the well‐known benefits of the object‐oriented development model, using objects to exchange information in JOPI is advantageous because it facilitates passing complex structures and enables the programmer to isolate the problem space from the parallelization problem. The run‐time environment for JOPI is portable, efficient and provides the necessary functionality to deploy and execute parallel Java programs. Experiments were conducted on a cluster system and a collection of heterogeneous platforms to measure JOPIs performance and compare it with MPI. The results show good performance gains using JOPI. Copyright