TensorFlow is a machine learning system that operates at large scale and in heterogeneous environments. Tensor-Flow uses dataflow graphs to represent computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general-purpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application developer: whereas in previous "parameter server" designs the management of shared state is built into the system, TensorFlow enables developers to experiment with novel optimizations and training algorithms. TensorFlow supports a variety of applications, with a focus on training and inference on deep neural networks. Several Google services use TensorFlow in production, we have released it as an open-source project, and it has become widely used for machine learning research. In this paper, we describe the TensorFlow dataflow model and demonstrate the compelling performance that TensorFlow achieves for several real-world applications.

/pdf/tensorflow-a-system-for-large-scale-machine-learning-10wgdjhl6k.pdf

TensorFlow: a system for large-scale machine learning

TensorFlow is an interface for expressing machine learning algorithms, and an
implementation for executing such algorithms. A computation expressed using
TensorFlow can be executed with little or no change on a wide variety of
heterogeneous systems, ranging from mobile devices such as phones and tablets
up to large-scale distributed systems of hundreds of machines and thousands of
computational devices such as GPU cards. The system is flexible and can be used
to express a wide variety of algorithms, including training and inference
algorithms for deep neural network models, and it has been used for conducting
research and for deploying machine learning systems into production across more
than a dozen areas of computer science and other fields, including speech
recognition, computer vision, robotics, information retrieval, natural language
processing, geographic information extraction, and computational drug
discovery. This paper describes the TensorFlow interface and an implementation
of that interface that we have built at Google. The TensorFlow API and a
reference implementation were released as an open-source package under the
Apache 2.0 license in November, 2015 and are available at www.tensorflow.org.

/pdf/tensorflow-large-scale-machine-learning-on-heterogeneous-2yk4fgchw6.pdf

TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems

TensorFlow is a machine learning system that operates at large scale and in heterogeneous environments. TensorFlow uses dataflow graphs to represent computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general-purpose GPUs, and custom designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application developer: whereas in previous "parameter server" designs the management of shared state is built into the system, TensorFlow enables developers to experiment with novel optimizations and training algorithms. TensorFlow supports a variety of applications, with particularly strong support for training and inference on deep neural networks. Several Google services use TensorFlow in production, we have released it as an open-source project, and it has become widely used for machine learning research. In this paper, we describe the TensorFlow dataflow model in contrast to existing systems, and demonstrate the compelling performance that TensorFlow achieves for several real-world applications.

TensorFlow: A system for large-scale machine learning

OceanStore is a utility infrastructure designed to span the globe and provide continuous access to persistent information. Since this infrastructure is comprised of untrusted servers, data is protected through redundancy and cryptographic techniques. To improve performance, data is allowed to be cached anywhere, anytime. Additionally, monitoring of usage patterns allows adaptation to regional outages and denial of service attacks; monitoring also enhances performance through pro-active movement of data. A prototype implementation is currently under development.

/pdf/oceanstore-an-architecture-for-global-scale-persistent-4qe9kbb44y.pdf

OceanStore: an architecture for global-scale persistent storage

The initial design of Apache Hadoop [1] was tightly focused on running massive, MapReduce jobs to process a web crawl. For increasingly diverse companies, Hadoop has become the data and computational agora---the de facto place where data and computational resources are shared and accessed. This broad adoption and ubiquitous usage has stretched the initial design well beyond its intended target, exposing two key shortcomings: 1) tight coupling of a specific programming model with the resource management infrastructure, forcing developers to abuse the MapReduce programming model, and 2) centralized handling of jobs' control flow, which resulted in endless scalability concerns for the scheduler. In this paper, we summarize the design, development, and current state of deployment of the next generation of Hadoop's compute platform: YARN. The new architecture we introduced decouples the programming model from the resource management infrastructure, and delegates many scheduling functions (e.g., task fault-tolerance) to per-application components. We provide experimental evidence demonstrating the improvements we made, confirm improved efficiency by reporting the experience of running YARN on production environments (including 100% of Yahoo! grids), and confirm the flexibility claims by discussing the porting of several programming frameworks onto YARN viz. Dryad, Giraph, Hoya, Hadoop MapReduce, REEF, Spark, Storm, Tez.

/pdf/apache-hadoop-yarn-yet-another-resource-negotiator-26ypcfjhgi.pdf

Apache Hadoop YARN: yet another resource negotiator

Google's Borg system is a cluster manager that runs hundreds of thousands of jobs, from many thousands of different applications, across a number of clusters each with up to tens of thousands of machines. It achieves high utilization by combining admission control, efficient task-packing, over-commitment, and machine sharing with process-level performance isolation. It supports high-availability applications with runtime features that minimize fault-recovery time, and scheduling policies that reduce the probability of correlated failures. Borg simplifies life for its users by offering a declarative job specification language, name service integration, real-time job monitoring, and tools to analyze and simulate system behavior. We present a summary of the Borg system architecture and features, important design decisions, a quantitative analysis of some of its policy decisions, and a qualitative examination of lessons learned from a decade of operational experience with it.

/pdf/large-scale-cluster-management-at-google-with-borg-4i7d9qnw5x.pdf

Large-scale cluster management at Google with Borg

Although disk storage densities are improving impressively (60% to 130% compounded annually), performance improvements have been occurring at only about 7% to 10% compounded annually over the last decade. As a result, disk system performance is fast becoming a dominant factor in overall svstem behavior. Naturally, researchers want to improve overall I/O performance, of which a large component is the performance of the disk drive itself. This research often involves using analytical or simulation models to compare alternative approaches, and the quality of these models determines the quality of the conclusions: indeed, the wrong modeling assumptions can lead to erroneous conclusions. Nevertheless, little work has been done to develop or describe accurate disk drive models. This may explain the commonplace use of simple, relatively inaccurate models. We believe there is much room for improvement. This article demonstrates and describes a calibrated, high-quality disk drive model in which the overall error factor is 14 times smaller than that of a simple first-order model. We describe the various disk drive performance components separately, then show how their inclusion improves the simulation model. This enables an informed trade-off between effort and accuracy. In addition, we provide detailed characteristics for two disk drives, as well as a brief description of a simulation environment that uses the disk drive model. >

/pdf/an-introduction-to-disk-drive-modeling-oetfqkqe10.pdf

An introduction to disk drive modeling

Increasing scale and the need for rapid response to changing requirements are hard to meet with current monolithic cluster scheduler architectures. This restricts the rate at which new features can be deployed, decreases efficiency and utilization, and will eventually limit cluster growth. We present a novel approach to address these needs using parallelism, shared state, and lock-free optimistic concurrency control.We compare this approach to existing cluster scheduler designs, evaluate how much interference between schedulers occurs and how much it matters in practice, present some techniques to alleviate it, and finally discuss a use case highlighting the advantages of our approach -- all driven by real-life Google production workloads.

Omega: flexible, scalable schedulers for large compute clusters

Elastic resource scaling lets cloud systems meet application service level objectives (SLOs) with minimum resource provisioning costs. In this paper, we present CloudScale, a system that automates fine-grained elastic resource scaling for multi-tenant cloud computing infrastructures. CloudScale employs online resource demand prediction and prediction error handling to achieve adaptive resource allocation without assuming any prior knowledge about the applications running inside the cloud. CloudScale can resolve scaling conflicts between applications using migration, and integrates dynamic CPU voltage/frequency scaling to achieve energy savings with minimal effect on application SLOs. We have implemented CloudScale on top of Xen and conducted extensive experiments using a set of CPU and memory intensive applications (RUBiS, Hadoop, IBM System S). The results show that CloudScale can achieve significantly higher SLO conformance than other alternatives with low resource and energy cost. CloudScale is non-intrusive and light-weight, and imposes negligible overhead (

/pdf/cloudscale-elastic-resource-scaling-for-multi-tenant-cloud-2vrynctts0.pdf

CloudScale: elastic resource scaling for multi-tenant cloud systems

Cloud systems require elastic resource allocation to minimize resource provisioning costs while meeting service level objectives (SLOs). In this paper, we present a novel PRedictive Elastic reSource Scaling (PRESS) scheme for cloud systems. PRESS unobtrusively extracts fine-grained dynamic patterns in application resource demands and adjust their resource allocations automatically. Our approach leverages light-weight signal processing and statistical learning algorithms to achieve online predictions of dynamic application resource requirements. We have implemented the PRESS system on Xen and tested it using RUBiS and an application load trace from Google. Our experiments show that we can achieve good resource prediction accuracy with less than 5% over-estimation error and near zero under-estimation error, and elastic resource scaling can both significantly reduce resource waste and SLO violations.

/pdf/press-predictive-elastic-resource-scaling-for-cloud-systems-4p6vqt5k2x.pdf

John Wilkes

Papers

Large-scale cluster management at Google with Borg

An introduction to disk drive modeling

Omega: flexible, scalable schedulers for large compute clusters

CloudScale: elastic resource scaling for multi-tenant cloud systems

PRESS: PRedictive Elastic ReSource Scaling for cloud systems