Benchmarking network fabrics for data distributed training of deep neural networks
MIT Lincoln Laboratory（麻省理工学院林肯实验室）
Siddharth Samsi, Andrew Prout, Michael Jones, Andrew Kirby, Bill Arcand, Bill Bergeron, David Bestor, Chansup Byun, Vijay Gadepally, Michael Houle, Matthew Hubb… 展开作者
Siddharth Samsi, Andrew Prout, Michael Jones, Andrew Kirby, Bill Arcand, Bill Bergeron, David Bestor, Chansup Byun, Vijay Gadepally, Michael Houle, Matthew Hubbell, Anna Klein, Peter Michaleas, Lauren Milechin, Julie Mullen, Antonio Rosa, Charles Yee, Albert Reuther, Jeremy Kepner
Artificial Intelligence/Machine Learning applications require the training of complex models on large amounts of labelled data. The large computational requirements for training deep models have necessitated the development of new methods for faster training. One such approach is the data parallel approach, where the training data is distributed across multiple compute nodes. This approach is simple to implement and supported by most of the commonly used machine learning frameworks. The data parallel approach leverages MPI for communicating gradients across all nodes. In this paper, we examine the effects of using different physical hardware interconnects and network-related software primitives for enabling data distributed deep learning. We compare the effect of using GPUDirect and NCCL on Ethernet and OmniPath fabrics. Our results show that using Ethernet-based networking in shared HPC systems does not have a significant effect on the training times for commonly used deep neural network architectures or traditional HPC applications such as Computational Fluid Dynamics.
We introduce Deep500: the first customizable benchmarking infrastructure that enables fair comparison of the plethora of deep learning frameworks, algorithms, libraries, and techniques. The key idea behind Deep500 is its modular design, where deep learning is factorized into four distinct levels: operators, network processing, training, and distributed training. Our evaluation illustrates that Deep500 is customizable (enables combining and benchmarking different deep learning codes) and fair (uses carefully selected metrics). Moreover, Deep500 is fast (incurs negligible overheads), verifiable …
University of Illinois, Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；IBM Thomas J. Watson Research Center（IBM托马斯·J·沃森研究中心）
Machine Learning (ML) and Deep Learning (DL) innovations are being introduced at such a rapid pace that researchers are hard-pressed to analyze and study them. The complicated procedures for evaluating innovations, along with the lack of standard and efficient ways of specifying and provisioning ML/DL evaluation, is a major "pain point" for the community. This paper proposes MLModelScope, an open-source, framework/hardware agnostic, extensible and customizable design that enables repeatable, fair, and scalable model evaluation and benchmarking. We implement the distributed design with support …
In the big data era of observational oceanography, passive acoustics datasets are becoming too high volume to be processed on local computers due to their processor and memory limitations. As a result there is a current need for our community to turn to cloud-based distributed computing. We present a scalable computing system for FFT (Fast Fourier Transform)-based features (e.g., Power Spectral Density) based on the Apache distributed frameworks Hadoop and Spark. These features are at the core of many different types of acoustic analysis where the need of processing data at scale with speed is…
Deep learning has been shown as a successful machine learning method for a variety of tasks, and its popularity results in numerous open-source deep learning software tools. Training a deep network is usually a very time-consuming process. To address the computational challenge in deep learning, many tools exploit hardware features such as multi-core CPUs and many-core GPUs to shorten the training time. However, different tools exhibit different features and running performance when training different types of deep networks on different hardware platforms, which makes it difficult for end user…
Benchmarking Robust Aggregation in Decentralized Gradient Marketplaces
Department of Computer Science and Engineering（计算机科学与工程系）；Pennsylvania State University（宾夕法尼亚州立大学）；Department of Computer Science（计算机科学系）；Cornell University（康奈尔大学）
The rise in popularity of permissioned blockchain platforms in recent time is significant. Hyperledger Fabric is one such permissioned blockchain platform and one of the Hyperledger projects hosted by the Linux Foundation. The Fabric comprises various components such as smart-contracts, endorsers, committers, validators, and orderers. As the performance of blockchain platform is a major concern for enterprise applications, in this work, we perform a comprehensive empirical study to characterize the performance of Hyperledger Fabric and identify potential performance bottlenecks to gain a bette…
Training and Hyperparameter Optimization (HPO) of deep learning-based AI models are often compute resource intensive and calls for the use of large-scale distributed resources as well as scalable and resource efficient hyperparameter search algorithms. This work studies the potential of using model performance prediction to aid the HPO process carried out on High Performance Computing systems. In addition, a quantum annealer is used to train the performance predictor and a method is proposed to overcome some of the problems derived from the current limitations in quantum systems as well as to …
Distributed stream processing engines are designed with a focus on scalability to process big data volumes in a continuous manner. We present the Theodolite method for benchmarking the scalability of distributed stream processing engines. Core of this method is the definition of use cases that microservices implementing stream processing have to fulfill. For each use case, our method identifies relevant workload dimensions that might affect the scalability of a use case. We propose to design one benchmark per use case and relevant workload dimension. We present a general benchmarking framework…
The rapid growth of blockchain systems leads to increasing interest in understanding and comparing blockchain performance at scale. In this paper, we focus on analyzing the performance of Hyperledger Fabric v1.1 - one of the most popular permissioned blockchain systems. Prior works have analyzed Hyperledger Fabric v0.6 in depth, but newer versions of the system undergo significant changes that warrant new analysis. Existing works on benchmarking the system are limited in their scope: some consider only small networks, others consider scalability of only parts of the system instead of the whole…
Fabric is a modular and extensible open-source system for deploying and operating permissioned blockchains and one of the Hyperledger projects hosted by the Linux Foundation (www.hyperledger.org). Fabric is the first truly extensible blockchain system for running distributed applications. It supports modular consensus protocols, which allows the system to be tailored to particular use cases and trust models. Fabric is also the first blockchain system that runs distributed applications written in standard, general-purpose programming languages, without systemic dependency on a native cryptocurr…
Deep learning (DL) workflows demand an ever-increasing budget of compute and energy in order to achieve outsized gains. Neural architecture searches, hyperparameter sweeps, and rapid prototyping consume immense resources that can prevent resource-constrained researchers from experimenting with large models and carry considerable environmental impact. As such, it becomes essential to understand how different deep neural networks (DNNs) and training leverage increasing compute and energy resources -- especially specialized computationally-intensive models across different domains and application…
Internet of Things (IoT) is a technology paradigm where millions of sensors monitor, and help inform or manage, physical, envi- ronmental and human systems in real-time. The inherent closed-loop re- sponsiveness and decision making of IoT applications makes them ideal candidates for using low latency and scalable stream processing plat- forms. Distributed Stream Processing Systems (DSPS) are becoming es- sential components of any IoT stack, but the efficacy and performance of contemporary DSPS have not been rigorously studied for IoT data streams and applications. Here, we develop a benchmark …
Modern high load applications store data using multiple database instances. Such an architecture requires data consistency, and it is important to ensure even distribution of data among nodes. Load balancing is used to achieve these goals. Hashing is the backbone of virtually all load balancing systems. Since the introduction of classic Consistent Hashing, many algorithms have been devised for this purpose. One of the purposes of the load balancer is to ensure storage cluster scalability. It is crucial for the performance of the whole system to transfer as few data records as possible during n…
Within the last few years, a countless number of blockchain systems have emerged on the market, each one claiming to revolutionize the way of distributed transaction processing in one way or the other. Many blockchain features, such as byzantine fault tolerance (BFT), are indeed valuable additions in modern environments. However, despite all the hype around the technology, many of the challenges that blockchain systems have to face are fundamental transaction management problems. These are largely shared with traditional database systems, which have been around for decades already. These simil…
Making serverless computing widely applicable requires detailed performance understanding. Although contemporary benchmarking approaches exist, they report only coarse results, do not apply distributed tracing, do not consider asynchronous applications, and provide limited capabilities for (root cause) analysis. Addressing this gap, we design and implement ServiBench, a serverless benchmarking suite. ServiBench (i) leverages synchronous and asynchronous serverless applications representative of production usage, (ii) extrapolates cloud-provider data to generate realistic workloads, (iii) condu…
Blockchain is one of the most popular distributed ledger technologies. It can solve the trust issue among enterprises. Hyperledger Fabric is a permissioned blockchain aiming at enterprise-grade business applications. However, compared to traditional distributed database solutions, one issue of blockchain based application development is the limited data access. For Fabric, the ledger data can only be retrieved by limited interfaces provided by Fabric SDKs or chaincode. In order to meet the requirements of data query and provide flexible query functions for real applications built on Fabric, th…
Hyperledger Fabric is a distributed operating system for permissioned blockchains hosted by the Linux Foundation. It is the first truly extensible blockchain system for running distributed applications at enterprise grade scale. To achieve this, Hyperledger Fabric introduces a novel execute-order-validate blockchain architecture, allowing parallelization of transaction execution and validation. However, this raises the need for transaction isolation. Today transaction isolation is attained by locking the entire state database during simulation of transactions and database updates. This lock is…
Infrastructures supporting distributed scientific collaborations must address competing goals in both providing high-performance access to resources while simultaneously securing the infrastructure against security threats. The NetBASILISK project is attempting to improve the security of such infrastructures while not adversely impacting their performance. This paper will present our work to create a benchmark and monitoring infrastructure that allows us to test for any degradation in transferring data into a NetBASILISK protected site.
Recent developments in blockchain technology have inspired innovative new designs in resilient distributed and database systems. At their core, these blockchain applications typically use Byzantine fault-tolerant consensus protocols to maintain a common state across all replicas, even if some replicas are faulty or malicious. Unfortunately, existing consensus protocols are not designed to deal with geo-scale deployments in which many replicas spread across a geographically large area participate in consensus. To address this, we present the Geo-Scale Byzantine FaultTolerant consensus protocol …
Hull is an accelerator-rich distributed implantable Brain-Computer Interface (BCI) that reads biological neurons at data rates that are 2-3 orders of magnitude higher than the prior state of art, while supporting many neuroscientific applications. Prior approaches have restricted brain interfacing to tens of megabits per second in order to meet two constraints necessary for effective operation and safe long-term implantation -- power dissipation under tens of milliwatts and response latencies in the tens of milliseconds. Hull also adheres to these constraints, but is able to interface with the…
SciDB is a scalable, computational database management system that uses an array model for data storage. The array data model of SciDB makes it ideally suited for storing and managing large amounts of imaging data. SciDB is designed to support advanced analytics in database, thus reducing the need for extracting data for analysis. It is designed to be massively parallel and can run on commodity hardware in a high performance computing (HPC) environment. In this paper, we present the performance of SciDB using simulated image data. The Dynamic Distributed Dimensional Data Model (D4M) software i…
The paper suggests a new approach based on blockchain technologies and smart contracts to creation of a distributed system for managing provenance metadata, as well as access rights to data in distributed storages, which is fault-tolerant, safe and secure from the point of view of preservation of metadata records from accidental or intentional distortions. The implementation of the proposed approach is based on the permissioned blockchains and on the Hyperledger Fabric blockchain platform in conjunction with Hyperledger Composer.
Blockchain, as a distributed ledger technology, becomes more and more popular in both industry and academia. Each peer in blockchain system maintains a copy of ledger and makes sure of data consistency through consensus protocol. Blockchain system can provide many benefits such as immutability, transparency and security. Hyperledger Fabric is permissioned blockchain platform hosted by Linux foundation. Fabric has various components such as peer, ordering service, chaincode and state database. The structure of Fabric network is very complicated to provide reliable permissioned blockchain servic…
Indiana University（印第安纳大学）；Argonne National Lab（阿贡国家实验室）
CNN-based surrogates have become prevalent in scientific applications to replace conventional time-consuming physical approaches. Although these surrogates can yield satisfactory results with significantly lower computation costs over small training datasets, our benchmarking results show that data-loading overhead becomes the major performance bottleneck when training surrogates with large datasets. In practice, surrogates are usually trained with high-resolution scientific data, which can easily reach the terabyte scale. Several state-of-the-art data loaders are proposed to improve the loadi…
Grid computing consists of the coordinated use of large sets of diverse, geographically distributed resources for high performance computation. Effective monitoring of these computing resources is extremely important to allow efficient use on the Grid. The large number of heterogeneous computing entities available in Grids makes the task challenging. In this work, we describe a Grid monitoring system, called GridMonitor, that captures and makes available the most important information from a large computing facility. The Grid monitoring system consists of four tiers: local monitoring, archivin…
The recent boom of big data, coupled with the challenges of its processing and storage gave rise to the development of distributed data processing and storage paradigms like MapReduce, Spark, and NoSQL databases. With the advent of cloud computing, processing and storing such massive datasets on clusters of machines is now feasible with ease. However, there are limited tools and approaches, which users can rely on to gauge and comprehend the performance of their big data applications deployed locally on clusters, or in the cloud. Researchers have started exploring this area by providing benchm…
In the fields of big data, AI, and streaming processing, we work with large amounts of data from multiple sources. Due to memory and network limitations, we process data streams on distributed systems to alleviate computational and network loads. When data streams with non-uniform distributions are processed, we often observe overloaded partitions due to the use of simple hash partitioning. To tackle this imbalance, we can use dynamic partitioning algorithms that require a sampling algorithm to precisely estimate the underlying distribution of the data stream. There is no standardized way to t…
Benchmarking and comparing performance of a scientific simulation across hardware platforms is a complex task. When the simulation in question is constructed with an asynchronous, many-task (AMT) runtime offloading work to GPUs, the task becomes even more complex. In this paper, we discuss the use of a uniquely suited performance measurement library, APEX, to capture the performance behavior of a simulation built on HPX, a highly scalable, distributed AMT runtime. We examine the performance of the astrophysics simulation carried-out by Octo-Tiger on two different supercomputing architectures. …
I describe the performance of the CRBLASTER computational framework on a 350-MHz 49-core Maestro Development Board (MDB). The 49-core Interim Test Chip (ITC) was developed by the U.S. Government and is based on the intellectual property of the 64-core TILE64 processor of the Tilera Corporation. The Maestro processor is intended for use in the high radiation environments found in space; the ITC was fabricated using IBM 90-nm CMOS 9SF technology and Radiation-Hardening-by-Design (RHDB) rules. CRBLASTER is a parallel-processing cosmic-ray rejection application based on a simple computational fram…
We propose TopoOpt, a novel direct-connect fabric for deep neural network (DNN) training workloads. TopoOpt co-optimizes the distributed training process across three dimensions: computation, communication, and network topology. We demonstrate the mutability of AllReduce traffic, and leverage this property to construct efficient network topologies for DNN training jobs. TopoOpt then uses an alternating optimization technique and a group theory-inspired algorithm called TotientPerms to find the best network topology and routing plan, together with a parallelization strategy. We build a fully fu…
