MapReduce is a programming model and an associated implementation for processing and generating large data sets. Users specify a map function that processes a key/value pair to generate a set of intermediate key/value pairs, and a reduce function that merges all intermediate values associated with the same intermediate key. Many real world tasks are expressible in this model, as shown in the paper. Programs written in this functional style are automatically parallelized and executed on a large cluster of commodity machines. The run-time system takes care of the details of partitioning the input data, scheduling the program’s execution across a set of machines, handling machine failures, and managing the required inter-machine communication. This allows programmers without any experience with parallel and distributed systems to easily utilize the resources of a large distributed system. Our implementation of MapReduce runs on a large cluster of commodity machines and is highly scalable: a typical MapReduce computation processes many terabytes of data on thousands of machines. Programmers find the system easy to use: hundreds of MapReduce programs have been implemented and upwards of one thousand MapReduce jobs are executed on Google’s clusters every day. 1

Very large data sets often have a flat but regular structure and span multiple disks and machines. Examples include telephone call records, network logs, and web document repositories. These large data sets are not amenable to study using traditional database techniques, if only because they can be too large to fit in a single relational database. On the other hand, many of the analyses done on them can be expressed using simple, easily distributed computations: filtering, aggregation, extraction of statistics, and so on. We present a system for automating such analyses. A filtering phase, in which a query is expressed using a new procedural programming language, emits data to an aggregation phase. Both phases are distributed over hundreds or even thousands of computers. The results are then collated and saved to a file. The design—including the separation into two phases, the form of the programming language, and the properties of the aggregators—exploits the parallelism inherent in having data and computation distributed across many machines. 1

Many scientific disciplines are now data and information driven, and new scientific knowledge is often gained by scientists putting together data analysis and knowledge discovery “pipelines”. A related trend is that more and more scientific communities realize the benefits of sharing their data and computational services, and are thus contributing to a distributed data and computational community infrastructure (a.k.a. “the Grid”). However, this infrastructure is only a means to an end and scientists ideally should be bothered little with its existence. The goal is for scientists to focus on development and use of what we call scientific workflows. These are networks of analytical steps that may involve, e.g., database access

To enable the rapid execution of many tasks on compute clusters, we have developed Falkon, a Fast and Light-weight tasK executiON framework. Falkon integrates (1) multi-level scheduling to separate resource acquisition (via, e.g., requests to batch schedulers) from task dispatch, and (2) a streamlined dispatcher. Falkon’s integration of multi-level scheduling and streamlined dispatchers delivers performance not provided by any other system. We describe Falkon architecture and implementation, and present performance results for both microbenchmarks and applications. Microbenchmarks show that Falkon throughput (487 tasks/sec) and scalability (to 54,000 executors and 2,000,000 tasks processed in just 112 minutes) are one to two orders of magnitude better than other systems used in production Grids. Large-scale astronomy and medical applications executed under Falkon by the Swift parallel programming system achieve up to 90 % reduction in end-to-end run time, relative to versions that execute tasks via separate scheduler submissions.

A shared distributed infrastructure is formed by federating computation resources from multiple domains. Such shared infrastructures are increasing in popularity and are providing massive amounts of aggregated computation resources to large numbers of users. Meanwhile, virtualization technologies, at machine and network levels, are maturing and enabling mutually isolated virtual computation environments for executing arbitrary parallel/distributed applications on top of such a shared physical infrastructure. In this paper, we go one step further by supporting autonomic adaptation of virtual computation environments as active, integrated entities. More specifically, driven by both dynamic availability of infrastructure resources and dynamic application resource demand, a virtual computation environment is able to automatically relocate itself across the infrastructure and scale its share of infrastructural resources. Such autonomic adaptation is transparent to both users of virtual environments and administrators of infrastructures, maintaining the look and feel of a stable, dedicated environment for the user. As our proofof-concept, we present the design, implementation, and evaluation of a system called VIOLIN, which is composed of a virtual network of virtual machines capable of live migration across a multi-domain physical infrastructure. 1

Among existing grid middleware approaches, one simple, powerful, and flexible approach consists of using servers available in different administrative domains through the classical client-server or Remote Procedure Call (RPC) paradigm. Network Enabled Servers implement this

As organizations start to use data-intensive cluster computing systems like Hadoop and Dryad for more applications, there is a growing need to share clusters between users. However, there is a conflict between fairness in scheduling and data locality (placing tasks on nodes that contain their input data). We illustrate this problem through our experience designing a fair scheduler for a 600-node Hadoop cluster at Facebook. To address the conflict between locality and fairness, we propose a simple algorithm called delay scheduling: when the job that should be scheduled next according to fairness cannot launch a local task, it waits for a small amount of time, letting other jobs launch tasks instead. We find that delay scheduling achieves nearly optimal data locality in a variety of workloads and can increase throughput by up to 2x while preserving fairness. In addition, the simplicity of delay scheduling makes it applicable under a wide variety of scheduling policies beyond fair sharing.

Ever more scientists are employing large-scale distributed systems such as grids for their computational work, instead of tightly coupled high-performance computing systems. However, while these distributed systems are more cost-effective, their heterogeneity in terms of hardware, software, and systems administration, and the lack of accurate resource information leads to inefficient scheduling. In addition, and in contrast to the workloads of tightly coupled high-performance computing systems, a large part of the workloads submitted to these distributed systems consists of large sets (bags) of sequential tasks. Therefore, a realistic performance analysis of scheduling bags-of-tasks in large-scale distributed systems is important. Towards this end, we introduce in this paper a realistic workload model for bags-of-tasks, and we explore through trace-based simulations the design space of scheduling bags-of-tasks. Finally, we identify three new scheduling policies that use only inaccurate information when scheduling, and we compare them against known classes of proposed scheduling policies.

Abstract—Cloud computing is an emerging commercial infrastructure paradigm that promises to eliminate the need for maintaining expensive computing facilities by companies and institutes alike.Throughtheuseofvirtualizationandresourcetime-sharing, clouds serve with a single set of physical resources a large user base withdifferentneeds.Thus,cloudshavethepotentialtoprovide to their owners the benefits of an economy of scale and, at the same time, becomeanalternativeforscientiststoclusters,grids,and parallel production environments. However, the current commercial clouds have been built to support web and small database workloads, which are very different from typical scientific computing workloads. Moreover, the use of virtualization and resource time-sharing may introduce significant performance penalties for the demanding scientific computing workloads. In this work we analyze the performance of cloud computing services for scientific computing workloads. We quantify the presence in real scientific computing workloads of Many-Task Computing (MTC) users, that is, of users who employ looselycoupledapplicationscomprisingmanytaskstoachieve their scientific goals. Then, we perform an empirical evaluation of theperformanceoffourcommercialcloudcomputingservices including Amazon EC2, which is currently the largest commercial cloud. Last,wecomparethroughtrace-basedsimulationtheperformance characteristics and cost models of clouds and other scientific computing platforms, for general and MTC-based scientific computing workloads. Our results indicate that the current clouds need an order of magnitude in performance improvement to be useful tothe scientific community, and show which improvements should be considered first to address this discrepancy between offer and demand.

Cloud Computing has become another buzzword after Web 2.0. However, there are dozens of different definitions for Cloud Computing and there seems to be no consensus on what a Cloud is. On the other hand, Cloud Computing is not a completely new concept; it has intricate connection to the relatively new but thirteen-year established Grid Computing paradigm, and other relevant technologies such as utility computing, cluster computing, and distributed systems in general. This paper strives to compare and contrast Cloud Computing with Grid Computing from various angles and give insights into the essential characteristics of both.