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

Covering the basic techniques used in the latest research work, the author consolidates progress made so far, including some very recent and promising results, and conveys the beauty and excitement of work in the field. He gives clear, lucid explanations of key results and ideas, with intuitive proofs, and provides critical examples and numerous illustrations to help elucidate the algorithms. Many of the results presented have been simplified and new insights provided. Of interest to theoretical computer scientists, operations researchers, and discrete mathematicians.

Approximation Algorithms

https://beloglazov.info/papers/2011-energy-aware-fgcs.pdf

Energy-aware resource allocation heuristics for efficient management of data centers for Cloud computing

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

Analysis of a Local Search Heuristic for Facility Location Problems

We describe the design of an agile data center with integrated server and storage virtualization technologies. Such data centers form a key building block for new cloud computing architectures. We also show how to leverage this integrated agility for non-disruptive load balancing in data centers across multiple resource layers - servers, switches, and storage. We propose a novel load balancing algorithm called VectorDot for handling the hierarchical and multi-dimensional resource constraints in such systems. The algorithm, inspired by the successful Toyoda method for multi-dimensional knapsacks, is the first of its kind.We evaluate our system on a range of synthetic and real data center testbeds comprising of VMware ESX servers, IBM SAN Volume Controller, Cisco and Brocade switches. Experiments under varied conditions demonstrate the end-to-end validity of our system and the ability of VectorDot to efficiently remove overloads on server, switch and storage nodes.

Server-storage virtualization: integration and load balancing in data centers

Virtualization technologies like VMware and Xen provide features to specify the minimum and maximum amount of resources that can be allocated to a virtual machine (VM) and a shares based mechanism for the hypervisor to distribute spare resources among contending VMs. However much of the existing work on VM placement and power consolidation in data centers fails to take advantage of these features. One of our experiments on a real testbed shows that leveraging such features can improve the overall utility of the data center by 47% or even higher. Motivated by these, we present a novel suite of techniques for placement and power consolidation of VMs in data centers taking advantage of the min-max and shares features inherent in virtualization technologies. Our techniques provide a smooth mechanism for power-performance tradeoffs in modern data centers running heterogeneous applications, wherein the amount of resources allocated to a VM can be adjusted based on available resources, power costs, and application utilities. We evaluate our techniques on a range of large synthetic data center setups and a small real data center testbed comprising of VMware ESX servers. Our experiments confirm the end-to-end validity of our approach and demonstrate that our final candidate algorithm, PowerExpandMinMax, consistently yields the best overall utility across a broad spectrum of inputs - varying VM sizes and utilities, varying server capacities and varying power costs - thus providing a practical solution for administrators.

Shares and utilities based power consolidation in virtualized server environments

Consider a hierarchical network in which each node periodically issues a request for an object drawn from a fixed set of unit-size objects. Suppose further that the following conditions are satisfied: the frequency with which each node accesses each object is known; each node has a cache of known capacity; any cache can be accessed by any node; and any request is satisfied by the closest node with a copy of the desired object, at a cost proportional to the distance between the accessing node and the closest copy. In such an environment, it is desirable to fill the available cache space with copies of objects in such a way that the average access cost is minimized. We provide both exact and approximate polynomial-time algorithms for this hierarchical placement problem. Our exact algorithm is based on a reduction to min-cost flow, and does not appear to be practical for large problem sizes. Thus we are motivated to search for a faster approximation algorithm. Our main result is a simple constant-factor approximation algorithm for the hierarchical placement problem that admits an efficient distributed implementation.

Madhukar R. Korupolu

Papers

Large-scale cluster management at Google with Borg

Analysis of a Local Search Heuristic for Facility Location Problems

Server-storage virtualization: integration and load balancing in data centers

Shares and utilities based power consolidation in virtualized server environments

Placement algorithms for hierarchical cooperative caching