Bogdan Ghit

Balanced resource allocations across multiple dynamic MapReduce clusters

Running multiple instances of the MapReduce framework concurrently in a multicluster system or datacenter enables data, failure, and version isolation, which is attractive for many organizations. It may also provide some form of performance isolation, but in order to achieve this in the face of time-varying workloads submitted to the MapReduce instances, a mechanism for dynamic resource (re-)allocations to those instances is required. In this paper, we present such a mechanism called Fawkes that attempts to balance the allocations to MapReduce instances so that they experience similar service levels. Fawkes proposes a new abstraction for deploying MapReduce instances on physical resources, the MR-cluster, which represents a set of resources that can grow and shrink, and that has a core on which MapReduce is installed with the usual data locality assumptions but that relaxes those assumptions for nodes outside the core. Fawkes dynamically grows and shrinks the active MR-clusters based on a family of weighting policies with weights derived from monitoring their operation. We empirically evaluate Fawkes on a multicluster system and show that it can deliver good performance and balanced resource allocations, even when the workloads of the MR-clusters are very uneven and bursty, with workloads composed from both synthetic and real-world benchmarks.

Resource Management for Dynamic MapReduce Clusters in Multicluster Systems

State-of-the-art MapReduce frameworks such as Hadoop can easily scale up to thousands of machines and to large numbers of users. Nevertheless, some users may require isolated environments to develop their applications and to process their data, which calls for multiple deployments of MR clusters within the same physical infrastructure. In this paper, we design and implement a resource management system to facilitate the on-demand isolated deployment of MapReduce clusters in multicluster systems. Deploying multiple MapReduce clusters enables four types of isolation, with respect to performance, to data management, to fault tolerance, and to versioning. To efficiently manage the underlying physical resources, we propose three provisioning policies for dynamically resizing MapReduce clusters, and we evaluate the performance of our system through experiments on a real multicluster.

The BTWorld use case for big data analytics: Description, MapReduce logical workflow, and empirical evaluation

The commoditization of big data analytics, that is, the deployment, tuning, and future development of big data processing platforms such as MapReduce, relies on a thorough understanding of relevant use cases and workloads. In this work we propose BTWorld, a use case for time-based big data analytics that is representative for processing data collected periodically from a global-scale distributed system. BTWorld enables a data-driven approach to understanding the evolution of BitTorrent, a global file-sharing network that has over 100 million users and accounts for a third of todays upstream traffic. We describe for this use case the analyst questions and the structure of a multi-terabyte data set. We design a MapReduce-based logical workflow, which includes three levels of data dependency - inter-query, inter-job, and intra-job - and a query diversity that make the BTWorld use case challenging for todays big data processing tools; the workflow can be instantiated in various ways in the MapReduce stack. Last, we instantiate this complex workflow using Pig-Hadoop-HDFS and evaluate the use case empirically. Our MapReduce use case has challenging features: small (kilobytes) to large (250 MB) data sizes per observed item, excellent (10-6) and very poor (102) selectivity, and short (seconds) to long (hours) job duration.

KOALA-C: A task allocator for integrated multicluster and multicloud environments

Companies, scientific communities, and individual scientists with varying requirements for their compute-intensive applications may want to use public Infrastructure-as-a-Service clouds to increase the capacity of the resources they have access to. To enable such access, resource managers that currently act as gateways to clusters may also do so for clouds, but for this they require new architectures and scheduling frameworks. In this paper, we present the design and implementation of KOALA-C, which is an extension of the KOALA multicluster scheduler to multicloud environments. KOALA-C enables uniform management across multicluster and multicloud environments by provisioning resources from both infrastructures and grouping them into clusters of resources called sites. KOALA-C incorporates a comprehensive list of policies for scheduling jobs across multiple (sets of) sites, including both traditional policies and two new policies inspired by the well-known TAGS task assignment policy in distributed-server systems. Finally, we evaluate KOALA-C through realistic simulations and real-world experiments, and show that the new architecture and in particular its new policies show promise in achieving good job slowdown with high resource utilization.

Scheduling Workloads of Workflows with Unknown Task Runtimes

Workflows are important computational tools in many branches of science, and because of the dependencies among their tasks and their widely different characteristics, scheduling them is a difficult problem. Most research on scheduling workflows has focused on the offline problem of minimizing the make span of single workflows with known task runtimes. The problem of scheduling multiple workflows has been addressed either in an offline fashion, or still with the assumption of known task runtimes. In this paper, we study the problem of scheduling workloads consisting of an arrival stream of workflows without task runtime estimates. The resource requirements of a workflow can significantly fluctuate during its execution. Thus, we present four scheduling policies for workloads of workflows with as their main feature the extent to which they reserve processors to workflows to deal with these fluctuations. We perform simulations with realistic synthetic workloads and we show that any form of processor reservation only decreases the overall system performance and that a greedy backfilling-like policy performs best.

Towards an Optimized Big Data Processing System

Scalable by design to very large computing systems such as grids and clouds, MapReduce is currently a major big data processing paradigm. Nevertheless, existing performance models for MapReduce only comply with specific workloads that process a small fraction of the entire data set, thus failing to assess the capabilities of the MapReduce paradigm under heavy workloads that process exponentially increasing data volumes. The goal of my PhD is to build and analyze a scalable and dynamic big data processing system, including storage (distributed file system), execution engine (MapReduce), and query language (Pig). My contributions for the first two years of PhD research are the following: 1) the design and implementation of a resource management system part of a MapReduce-based processing system for deploying and resizing MapReduce clusters over multicluster systems, 2) the design and implementation of a benchmarking tool for the MapReduce processing system, and 3) the evaluation and modeling of MapReduce using workloads with very large data sets. Furthermore, based on the first two years research, we will optimize the MapReduce system to efficiently process terabytes of data.

Tyrex: Size-Based Resource Allocation in MapReduce Frameworks

Many large-scale data analytics infrastructures are employed for a wide variety of jobs, ranging from short interactive queries to large data analysis jobs that may take hours or even days to complete. As a consequence, data-processing frameworks like MapReduce may have workloads consisting of jobs with heavy-tailed processing requirements. With such workloads, short jobs may experience slowdowns that are an order of magnitude larger than large jobs do, while the users may expect slowdowns that are more in proportion with the job sizes. To address this problem of large job slowdown variability in MapReduce frameworks, we design a scheduling system called TYREX that is inspired by the well-known TAGS task assignment policy in distributed-server systems. In particular, TYREX partitions the resources of a MapReduce framework, allowing anyjob running in any partition to read data stored on any machine, imposes runtime limits in the partitions, and successively executesparts of jobs in a work-conserving way in these partitions untilthey can run to completion. We develop a statistical modelfor dynamically setting the runtime limits that achieves nearoptimaljob slowdown performance, and we empirically evaluate TYREX on a cluster system with workloads consisting of both synthetic and real-world benchmarks. We find that TYREX cuts in half the job slowdown variability while preserving the median job slowdown when compared to state-of-the-art MapReduce schedulers such as FIFO and FAIR. Furthermore, TYREX reduces the job slowdown at the 95th percentile by more than 50% when compared to FIFO and by 20-40% when compared to FAIR.

V for Vicissitude: The Challenge of Scaling Complex Big Data Workflows

In this paper we present the scaling of BTWorld, our MapReduce-based approach to observing and analyzing the global BitTorrent network which we have been monitoring for the past 4 years. BTWorld currently provides a comprehensive and complex set of queries implemented in Pig Latin, with data dependencies between them, which translate to several MapReduce jobs that have a heavy-tailed distribution with respect to both execution time and input size characteristics. Processing BitTorrent data in excess of 1 TB with our BTWorld workflow required an in-depth analysis of the entire software stack and the design of a complete optimization cycle. We analyze our system from both theoretical and experimental perspectives and we show how we attained a 15 times larger scale of data processing than our previous results.

Reducing Job Slowdown Variability for Data-Intensive Workloads

A well-known problem when executing data-intensive workloads with such frameworks as MapReduce is that small jobs with processing requirements counted in the minutes may suffer from the presence of huge jobs requiring hours or days of compute time, leading to a job slowdown distribution that is very variable and that is uneven across jobs of different sizes. Previous solutions to this problem for sequential or rigid jobs in single-server and distributed-server systems include priority-based FeedBack Queueing (FBQ), and Task Assignment by Guessing Sizes (TAGS), which kills and restarts from scratch on another server jobs that exceed the local time limit. In this paper, we derive four scheduling policies that are rightful descendants of existing size-based scheduling disciplines (among which FBQ and TAGS) with appropriate adaptations to data-intensive frameworks. The two main mechanisms employed by these policies are partitioning the resources of the datacenter, and isolating jobs with different size ranges. We evaluate these policies by means of realistic simulations of representative MapReduce workloads from Facebook and show that under the best of these policies, the vast majority of short jobs in MapReduce workloads experience close to ideal job slowdowns even under high system loads (in the range of 0.7-0.9) while the slowdown of the very large jobs is not prohibitive. We validate our simulations by means of experiments on a real multicluster system, and we find that the job slowdown performance results obtained with both match remarkably well.

Better Safe than Sorry: Grappling with Failures of In-Memory Data Analytics Frameworks

Providing fault-tolerance is of major importance for data analytics frameworks such as Hadoop and Spark, which are typically deployed in large clusters that are known to experience high failures rates. Unexpected events such as compute node failures are in particular an important challenge for in-memory data analytics frameworks, as the widely adopted approach to deal with them is to recompute work already done. Recomputing lost work, however, requires allocation of extra resource to re-execute tasks, thus increasing the job runtimes. To address this problem, we design a checkpointing system called Panda that is tailored to the intrinsic characteristics of data analytics frameworks. In particular, Panda employs fine-grained checkpointing at the level of task outputs and dynamically identifies tasks that are worthwhile to be checkpointed rather than be recomputed. As has been abundantly shown, tasks of data analytics jobs may have very variable runtimes and output sizes. These properties form the basis of three checkpointing policies which we incorporate into Panda. We first empirically evaluate Panda on a multicluster system with single data analytics applications under space-correlated failures, and find that Panda is close to the performance of a fail-free execution in unmodified Spark for a large range of concurrent failures. Then we perform simulations of complete workloads, mimicking the size and operation of a Google cluster, and show that Panda provides significant improvements in the average job runtime for wide ranges of the failure rate and system load.