THOUGHTS ABOUT THE FUTURE OF I/O

Size: px
Start display at page:

Download "THOUGHTS ABOUT THE FUTURE OF I/O"

Transcription

1 THOUGHTS ABOUT THE FUTURE OF I/O Dagstuhl Seminar Challenges and Opportunities of User-Level File Systems for HPC Franz-Josef Pfreundt, May 2017 Deep Learning I/O Challenges Memory Centric Computing : The Machine Low latency Non Volatile Memory Fraunhofer ITWM 2017 pfreundt 1

2 Fraunhofer ITWM HPC Department Research and Development of Industry Applications Parallel programming models Parallel file system Large scale visualization IoT in the energy market Development of parallel industry applications Performance Engineering Fraunhofer ITWM 2017 pfreundt 2

3 The one slide about BeeGFS Scalable IOPS Excellent N:1 performance, shared file I/O IOPS (Random 4k writes) up to 20 servers, 160 client procs Sequential I/O, 1 shared file, 600k block size up to 20 servers, 192 client procs IOPS MB/s Write Read # Storage servers # Servers Server components are in user space, client in kernel space X86 - ARM -POWER Low latency implementation No depency on Linux Kernel or Linux Distribution, any local FS ( ZFS, EXT, XFS, BTRFS, tmpfs) Very efficient multithreaded implementation > hyperconverged solution Fraunhofer ITWM 2017 pfreundt 3

4 The slide about BeeGFS on Demand - BeeOND = burst buffer Cray CS400 at Alfred Wegner Institut Broadwell CPU Omnipath Interconnect 0,5 TB SSD in each node BeeOND IOR 50TB Stripe size 1, local Stripe 4 Stripe size 1, any 308 Nodes write 160 GB/sec 161 GB/sec 160 GB/sec 308 Nodes read 167 GB/sec 164 GB/sec 167 GB/sec TSUBAME 3.0 plans to run the CN attached NVMe with BeeOND on 1 PByte of NVMe Fraunhofer ITWM 2017 pfreundt 4

5 Deep Learning I/O Challenge ( example Imagenet) Single Node I/O into a Lustre PFS ( Single GPU, FDR IB) IBM Minsky : 4 P100, NVLink Needs multiple SSD s in Raid 0 To allow sclability across 4 GPU s During training the data has be read 100 times ( 120 Mio file reads using standard Caffe) YFCC100m - a new public data set for multi media research 2D 2D+time By Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, Li- Jia Li, Communications of the ACM, Vol. 59 No. 2, Pages ,206,564 photos and 793,436 videos from 581,099 different photographers Size : About15 TByte Fraunhofer ITWM 2017 pfreundt 5

6 Our Solution BeeOND Build temporary parallel file system across nodes ( on demand per user) using BeeOND Every compute node can become a MDS Combine all files in one large binary with fixed offsets Or : all in memory ( rewrite the I/O layer in the DL framework) Fraunhofer ITWM 2017 pfreundt 6

7 Data Management : One key problem in parallel computing Its too complicated for application developers CPU We usually see nasty I/O patterns and bad performance Caches High Bandwidth Memory DRAM High Bandwidth low latency communication Non Volatile Memory, µs Latency Flash Storage Spinning Discs, Tape (Parallel) File I/O Fraunhofer ITWM 2017 pfreundt 7

8 GPI-2 Global Address Space Communication Interface Partitioned global address space Explicit one-sided communication with notification Every thread can communicate Multiple memory segments, Zero copy data transfer Standardized API ( GASPI) Developed and used at Fraunhofer since 2006 Complete replacement for MPI in industry applications GPLv3 Fraunhofer ITWM 2017 pfreundt 8

9 GPI Global Address Space Programming Interface 2) Hide latency by asynchronous one sided communication : RDMA 3. Every CPU Core can communicate and does not spent cycles for communication 1) Map pinned memory in a global address space Fraunhofer ITWM 2017 pfreundt 9

10 GPI-Space : Our approach to memory centric computing ( 2009) 1. Memory Virtualization using GPI 2. Concurrency and task management Virtual Global Memory Interconnect Application independend memory space Can keep data without an application running Applications are local to a node - Tasks Data exchange between Application and VMEM through a shared memory segment Data transfer between nodes with GPI Allows to couple tasks written in different languages Carl Adam Petri 1962: Description language for asynchronous and concurrent systems in order to add resources to running jobs simple, graphical representation (physical) properties: locality (no global state) concurrency (no total order given, just data dependencies) reversibility (calculate cause from effect) based on states not events (separate activation from execution) + some extensions ( names ports, type safety, expressions..) Fraunhofer ITWM 2017 pfreundt 10

11 GPI-Space is becomming a distributed OS 3. DRTS : Distributed Runtime System Debugging by on the fly modification of the Petri net Step Failure tolerant JIT compilation and execution of the Petri net Resources have capabilities : GPU, CPI, I/O Coscheduling of multi-node tasks (MPI) Preemptive scheduling of data transfers ( if information provided by the task) Fraunhofer ITWM 2017 pfreundt 11

12 GPI-Space + Domain Knowledge Complile high level workflows into Petri nets Dataflow modell in Seismic data processing Fraunhofer ITWM 2017 pfreundt 12

13 Example SPLOTCH : Visualization in Astrophysics LRZ Munich MPI Programm with problems Rewrite in GPI-Space in 3 weeks Thoughput : time to solution 10 x Fraunhofer ITWM 2017 pfreundt 13

14 Deep Learning on demand The development of new DNN s requires a lot of test runs - How can I do this cheap? Auto scaling Fail save: auto recovery, restarting Exploit the AWS spot market Automatic meta-parameter search Automatic data-management Supports original DL model descriptors e.g. Caffe & Tensor Flow Arbitrary Hardware nodes: GPU, CPU Developed in a few weeks Fraunhofer ITWM 2017 pfreundt 14

15 Deep Learning on Demand - Architecture Fraunhofer ITWM 2017 pfreundt 15

16 Data Management : One key problem in parallel computing Its too complicated for application developers CPU We usually see nasty I/O patterns and bad performance Caches High Bandwidth Memory DRAM High Bandwidth low latency communication Non Volatile Memory, µs Latency Flash Storage Spinning Discs, Tape (Parallel) File I/O Fraunhofer ITWM 2017 pfreundt 16

17 Directory/Cache API to support VMEM Multilevel Abstract Data Representation Allocation and global range Server knowledge Logical segment Segment knowledge Physical segment Physical segment: segment type and hardware dependent distribution Logical segment: linear view on physical segment Allocation: linear view on (distributed) part(s) of a segment Global range: subrange of an allocation The directory/cache unifies access to segments and abstracts distributed hardware no knowledge about data dependencies no knowledge about the runtime system behavior Fraunhofer ITWM 2017 pfreundt 17

18 VMEM Directory/Cache : Client Server Architecture The original data is stored in one or more segments across several nodes. Copies of global memory regions are stored in local caches. A local server may create and manage multiple local caches. Multiple clients may share local caches. External programs can connect to an already running directory/cache service. Tolerant to client failures (provided the clients are started in different processes) Fraunhofer ITWM 2017 pfreundt 18

19 Goal : Support the task based runtimes with an open source implementation OmpSs Runtime StarPU Runtime GPI-Space Directory/Cache API GASPI Segment MPI Segment BeeGFS/BeeOND Segment Keep some data in non-volatile memory Automate data transfer from storage to memory The API provides functions that may be used for taking scheduling decisions: transfer costs associated with a list of operations data locality information. The VMEM will become non volatile and data survive the appliaction Fraunhofer ITWM 2017 pfreundt 19

20 Moving on to byte addressable SCM Legacy Code _MPI GPI-Space Task World App App App App App App App POSIX I/O BeeGFS Client MD Server MD Server Key-Value Store VMEM/ API Storage Server Translate Posix Into Memory Operation Storage Server PC Cluster FDR IB SCM, PGAS Object Storage Fraunhofer ITWM 2017 pfreundt 20

21 Questions? Our plan until 2021 J. Keuper at Rice O&G Conference "Scaling Deep Learning Applications" Fraunhofer ITWM 2017 pfreundt 21

BeeGFS. Parallel Cluster File System. Container Workshop ISC July Marco Merkel VP ww Sales, Consulting

BeeGFS.   Parallel Cluster File System. Container Workshop ISC July Marco Merkel VP ww Sales, Consulting BeeGFS The Parallel Cluster File System Container Workshop ISC 28.7.18 www.beegfs.io July 2018 Marco Merkel VP ww Sales, Consulting HPC & Cognitive Workloads Demand Today Flash Storage HDD Storage Shingled

More information

CafeGPI. Single-Sided Communication for Scalable Deep Learning

CafeGPI. Single-Sided Communication for Scalable Deep Learning CafeGPI Single-Sided Communication for Scalable Deep Learning Janis Keuper itwm.fraunhofer.de/ml Competence Center High Performance Computing Fraunhofer ITWM, Kaiserslautern, Germany Deep Neural Networks

More information

Asynchronous Parallel Stochastic Gradient Descent. A Numeric Core for Scalable Distributed Machine Learning Algorithms

Asynchronous Parallel Stochastic Gradient Descent. A Numeric Core for Scalable Distributed Machine Learning Algorithms Asynchronous Parallel Stochastic Gradient Descent A Numeric Core for Scalable Distributed Machine Learning Algorithms J. Keuper and F.-J. Pfreundt Competence Center High Performance Computing Fraunhofer

More information

Distributed Training of Deep Neural Networks: Theoretical and Practical Limits of Parallel Scalability

Distributed Training of Deep Neural Networks: Theoretical and Practical Limits of Parallel Scalability Distributed Training of Deep Neural Networks: Theoretical and Practical Limits of Parallel Scalability Janis Keuper Itwm.fraunhofer.de/ml Competence Center High Performance Computing Fraunhofer ITWM, Kaiserslautern,

More information

Technologies for High Performance Data Analytics

Technologies for High Performance Data Analytics Technologies for High Performance Data Analytics Dr. Jens Krüger Fraunhofer ITWM 1 Fraunhofer ITWM n Institute for Industrial Mathematics n Located in Kaiserslautern, Germany n Staff: ~ 240 employees +

More information

Towards Scalable Machine Learning

Towards Scalable Machine Learning Towards Scalable Machine Learning Janis Keuper itwm.fraunhofer.de/ml Competence Center High Performance Computing Fraunhofer ITWM, Kaiserslautern, Germany Fraunhofer Center Machnine Larning Outline I Introduction

More information

Introducing the Cray XMT. Petr Konecny May 4 th 2007

Introducing the Cray XMT. Petr Konecny May 4 th 2007 Introducing the Cray XMT Petr Konecny May 4 th 2007 Agenda Origins of the Cray XMT Cray XMT system architecture Cray XT infrastructure Cray Threadstorm processor Shared memory programming model Benefits/drawbacks/solutions

More information

Application Example Running on Top of GPI-Space Integrating D/C

Application Example Running on Top of GPI-Space Integrating D/C Application Example Running on Top of GPI-Space Integrating D/C Tiberiu Rotaru Fraunhofer ITWM This project is funded from the European Union s Horizon 2020 Research and Innovation programme under Grant

More information

An Introduction to BeeGFS

An Introduction to BeeGFS An Introduction to BeeGFS Solid, fast, flexible and easy! www.beegfs.com Des données au BigData 13.12.2016 Bernd Lietzow An Introduction to BeeGFS Introduction BeeGFS Architecture BeeOND BeeGFS on Demand

More information

Feedback on BeeGFS. A Parallel File System for High Performance Computing

Feedback on BeeGFS. A Parallel File System for High Performance Computing Feedback on BeeGFS A Parallel File System for High Performance Computing Philippe Dos Santos et Georges Raseev FR 2764 Fédération de Recherche LUmière MATière December 13 2016 LOGO CNRS LOGO IO December

More information

Analyzing I/O Performance on a NEXTGenIO Class System

Analyzing I/O Performance on a NEXTGenIO Class System Analyzing I/O Performance on a NEXTGenIO Class System holger.brunst@tu-dresden.de ZIH, Technische Universität Dresden LUG17, Indiana University, June 2 nd 2017 NEXTGenIO Fact Sheet Project Research & Innovation

More information

DDN and Flash GRIDScaler, Flashscale Infinite Memory Engine

DDN and Flash GRIDScaler, Flashscale Infinite Memory Engine 1! DDN and Flash GRIDScaler, Flashscale Infinite Memory Engine T. Cecchi - September 21 st 2016 HPC Advisory Council 2! DDN END-TO-END DATA LIFECYCLE MANAGEMENT BURST & COMPUTE SSD, DISK & FILE SYSTEM

More information

HPC Storage Use Cases & Future Trends

HPC Storage Use Cases & Future Trends Oct, 2014 HPC Storage Use Cases & Future Trends Massively-Scalable Platforms and Solutions Engineered for the Big Data and Cloud Era Atul Vidwansa Email: atul@ DDN About Us DDN is a Leader in Massively

More information

Data Movement & Tiering with DMF 7

Data Movement & Tiering with DMF 7 Data Movement & Tiering with DMF 7 Kirill Malkin Director of Engineering April 2019 Why Move or Tier Data? We wish we could keep everything in DRAM, but It s volatile It s expensive Data in Memory 2 Why

More information

Aerie: Flexible File-System Interfaces to Storage-Class Memory [Eurosys 2014] Operating System Design Yongju Song

Aerie: Flexible File-System Interfaces to Storage-Class Memory [Eurosys 2014] Operating System Design Yongju Song Aerie: Flexible File-System Interfaces to Storage-Class Memory [Eurosys 2014] Operating System Design Yongju Song Outline 1. Storage-Class Memory (SCM) 2. Motivation 3. Design of Aerie 4. File System Features

More information

High-Performance Lustre with Maximum Data Assurance

High-Performance Lustre with Maximum Data Assurance High-Performance Lustre with Maximum Data Assurance Silicon Graphics International Corp. 900 North McCarthy Blvd. Milpitas, CA 95035 Disclaimer and Copyright Notice The information presented here is meant

More information

ZEST Snapshot Service. A Highly Parallel Production File System by the PSC Advanced Systems Group Pittsburgh Supercomputing Center 1

ZEST Snapshot Service. A Highly Parallel Production File System by the PSC Advanced Systems Group Pittsburgh Supercomputing Center 1 ZEST Snapshot Service A Highly Parallel Production File System by the PSC Advanced Systems Group Pittsburgh Supercomputing Center 1 Design Motivation To optimize science utilization of the machine Maximize

More information

Improved Solutions for I/O Provisioning and Application Acceleration

Improved Solutions for I/O Provisioning and Application Acceleration 1 Improved Solutions for I/O Provisioning and Application Acceleration August 11, 2015 Jeff Sisilli Sr. Director Product Marketing jsisilli@ddn.com 2 Why Burst Buffer? The Supercomputing Tug-of-War A supercomputer

More information

IME (Infinite Memory Engine) Extreme Application Acceleration & Highly Efficient I/O Provisioning

IME (Infinite Memory Engine) Extreme Application Acceleration & Highly Efficient I/O Provisioning IME (Infinite Memory Engine) Extreme Application Acceleration & Highly Efficient I/O Provisioning September 22 nd 2015 Tommaso Cecchi 2 What is IME? This breakthrough, software defined storage application

More information

Parallel Stochastic Gradient Descent: The case for native GPU-side GPI

Parallel Stochastic Gradient Descent: The case for native GPU-side GPI Parallel Stochastic Gradient Descent: The case for native GPU-side GPI J. Keuper Competence Center High Performance Computing Fraunhofer ITWM, Kaiserslautern, Germany Mark Silberstein Accelerated Computer

More information

An introduction to BeeGFS. Frank Herold, Sven Breuner June 2018 v2.0

An introduction to BeeGFS. Frank Herold, Sven Breuner June 2018 v2.0 An introduction to BeeGFS Frank Herold, Sven Breuner June 2018 v2.0 Abstract The scope of this paper is to give an overview on the parallel cluster file system BeeGFS and its basic concepts. It is intended

More information

CS370 Operating Systems

CS370 Operating Systems CS370 Operating Systems Colorado State University Yashwant K Malaiya Spring 2018 Lecture 22 File Systems Slides based on Text by Silberschatz, Galvin, Gagne Various sources 1 1 Disk Structure Disk can

More information

Results from TSUBAME3.0 A 47 AI- PFLOPS System for HPC & AI Convergence

Results from TSUBAME3.0 A 47 AI- PFLOPS System for HPC & AI Convergence Results from TSUBAME3.0 A 47 AI- PFLOPS System for HPC & AI Convergence Jens Domke Research Staff at MATSUOKA Laboratory GSIC, Tokyo Institute of Technology, Japan Omni-Path User Group 2017/11/14 Denver,

More information

CHAPTER 11: IMPLEMENTING FILE SYSTEMS (COMPACT) By I-Chen Lin Textbook: Operating System Concepts 9th Ed.

CHAPTER 11: IMPLEMENTING FILE SYSTEMS (COMPACT) By I-Chen Lin Textbook: Operating System Concepts 9th Ed. CHAPTER 11: IMPLEMENTING FILE SYSTEMS (COMPACT) By I-Chen Lin Textbook: Operating System Concepts 9th Ed. File-System Structure File structure Logical storage unit Collection of related information File

More information

Toward An Integrated Cluster File System

Toward An Integrated Cluster File System Toward An Integrated Cluster File System Adrien Lebre February 1 st, 2008 XtreemOS IP project is funded by the European Commission under contract IST-FP6-033576 Outline Context Kerrighed and root file

More information

DDN About Us Solving Large Enterprise and Web Scale Challenges

DDN About Us Solving Large Enterprise and Web Scale Challenges 1 DDN About Us Solving Large Enterprise and Web Scale Challenges History Founded in 98 World s Largest Private Storage Company Growing, Profitable, Self Funded Headquarters: Santa Clara and Chatsworth,

More information

I/O and Scheduling aspects in DEEP-EST

I/O and Scheduling aspects in DEEP-EST I/O and Scheduling aspects in DEEP-EST Norbert Eicker Jülich Supercomputing Centre & University of Wuppertal The research leading to these results has received funding from the European Community's Seventh

More information

NEXTGenIO Performance Tools for In-Memory I/O

NEXTGenIO Performance Tools for In-Memory I/O NEXTGenIO Performance Tools for In- I/O holger.brunst@tu-dresden.de ZIH, Technische Universität Dresden 22 nd -23 rd March 2017 Credits Intro slides by Adrian Jackson (EPCC) A new hierarchy New non-volatile

More information

Data Management. Parallel Filesystems. Dr David Henty HPC Training and Support

Data Management. Parallel Filesystems. Dr David Henty HPC Training and Support Data Management Dr David Henty HPC Training and Support d.henty@epcc.ed.ac.uk +44 131 650 5960 Overview Lecture will cover Why is IO difficult Why is parallel IO even worse Lustre GPFS Performance on ARCHER

More information

Oncilla - a Managed GAS Runtime for Accelerating Data Warehousing Queries

Oncilla - a Managed GAS Runtime for Accelerating Data Warehousing Queries Oncilla - a Managed GAS Runtime for Accelerating Data Warehousing Queries Jeffrey Young, Alex Merritt, Se Hoon Shon Advisor: Sudhakar Yalamanchili 4/16/13 Sponsors: Intel, NVIDIA, NSF 2 The Problem Big

More information

A ClusterStor update. Torben Kling Petersen, PhD. Principal Architect, HPC

A ClusterStor update. Torben Kling Petersen, PhD. Principal Architect, HPC A ClusterStor update Torben Kling Petersen, PhD Principal Architect, HPC Sonexion (ClusterStor) STILL the fastest file system on the planet!!!! Total system throughput in excess on 1.1 TB/s!! 2 Software

More information

The Leading Parallel Cluster File System

The Leading Parallel Cluster File System The Leading Parallel Cluster File System www.thinkparq.com www.beegfs.io ABOUT BEEGFS What is BeeGFS BeeGFS (formerly FhGFS) is the leading parallel cluster file system, developed with a strong focus on

More information

IBM CORAL HPC System Solution

IBM CORAL HPC System Solution IBM CORAL HPC System Solution HPC and HPDA towards Cognitive, AI and Deep Learning Deep Learning AI / Deep Learning Strategy for Power Power AI Platform High Performance Data Analytics Big Data Strategy

More information

Next-Generation NVMe-Native Parallel Filesystem for Accelerating HPC Workloads

Next-Generation NVMe-Native Parallel Filesystem for Accelerating HPC Workloads Next-Generation NVMe-Native Parallel Filesystem for Accelerating HPC Workloads Liran Zvibel CEO, Co-founder WekaIO @liranzvibel 1 WekaIO Matrix: Full-featured and Flexible Public or Private S3 Compatible

More information

Harmonia: An Interference-Aware Dynamic I/O Scheduler for Shared Non-Volatile Burst Buffers

Harmonia: An Interference-Aware Dynamic I/O Scheduler for Shared Non-Volatile Burst Buffers I/O Harmonia Harmonia: An Interference-Aware Dynamic I/O Scheduler for Shared Non-Volatile Burst Buffers Cluster 18 Belfast, UK September 12 th, 2018 Anthony Kougkas, Hariharan Devarajan, Xian-He Sun,

More information

FVM - How to program the Multi-Core FVM instead of MPI

FVM - How to program the Multi-Core FVM instead of MPI FVM - How to program the Multi-Core FVM instead of MPI DLR, 15. October 2009 Dr. Mirko Rahn Competence Center High Performance Computing and Visualization Fraunhofer Institut for Industrial Mathematics

More information

Virtual File System -Uniform interface for the OS to see different file systems.

Virtual File System -Uniform interface for the OS to see different file systems. Virtual File System -Uniform interface for the OS to see different file systems. Temporary File Systems -Disks built in volatile storage NFS -file system addressed over network File Allocation -Contiguous

More information

On the Use of Burst Buffers for Accelerating Data-Intensive Scientific Workflows

On the Use of Burst Buffers for Accelerating Data-Intensive Scientific Workflows On the Use of Burst Buffers for Accelerating Data-Intensive Scientific Workflows Rafael Ferreira da Silva, Scott Callaghan, Ewa Deelman 12 th Workflows in Support of Large-Scale Science (WORKS) SuperComputing

More information

Deep Learning on SHARCNET:

Deep Learning on SHARCNET: Deep Learning on SHARCNET: Best Practices Fei Mao Outlines What does SHARCNET have? - Hardware/software resources now and future How to run a job? - A torch7 example How to train in parallel: - A Theano-based

More information

Application Performance on IME

Application Performance on IME Application Performance on IME Toine Beckers, DDN Marco Grossi, ICHEC Burst Buffer Designs Introduce fast buffer layer Layer between memory and persistent storage Pre-stage application data Buffer writes

More information

Technical Computing Suite supporting the hybrid system

Technical Computing Suite supporting the hybrid system Technical Computing Suite supporting the hybrid system Supercomputer PRIMEHPC FX10 PRIMERGY x86 cluster Hybrid System Configuration Supercomputer PRIMEHPC FX10 PRIMERGY x86 cluster 6D mesh/torus Interconnect

More information

An Overview of Fujitsu s Lustre Based File System

An Overview of Fujitsu s Lustre Based File System An Overview of Fujitsu s Lustre Based File System Shinji Sumimoto Fujitsu Limited Apr.12 2011 For Maximizing CPU Utilization by Minimizing File IO Overhead Outline Target System Overview Goals of Fujitsu

More information

HPC Architectures. Types of resource currently in use

HPC Architectures. Types of resource currently in use HPC Architectures Types of resource currently in use Reusing this material This work is licensed under a Creative Commons Attribution- NonCommercial-ShareAlike 4.0 International License. http://creativecommons.org/licenses/by-nc-sa/4.0/deed.en_us

More information

Optimizing Out-of-Core Nearest Neighbor Problems on Multi-GPU Systems Using NVLink

Optimizing Out-of-Core Nearest Neighbor Problems on Multi-GPU Systems Using NVLink Optimizing Out-of-Core Nearest Neighbor Problems on Multi-GPU Systems Using NVLink Rajesh Bordawekar IBM T. J. Watson Research Center bordaw@us.ibm.com Pidad D Souza IBM Systems pidsouza@in.ibm.com 1 Outline

More information

Deep Learning mit PowerAI - Ein Überblick

Deep Learning mit PowerAI - Ein Überblick Stephen Lutz Deep Learning mit PowerAI - Open Group Master Certified IT Specialist Technical Sales IBM Cognitive Infrastructure IBM Germany Ein Überblick Stephen.Lutz@de.ibm.com What s that? and what s

More information

Toward portable I/O performance by leveraging system abstractions of deep memory and interconnect hierarchies

Toward portable I/O performance by leveraging system abstractions of deep memory and interconnect hierarchies Toward portable I/O performance by leveraging system abstractions of deep memory and interconnect hierarchies François Tessier, Venkatram Vishwanath, Paul Gressier Argonne National Laboratory, USA Wednesday

More information

DELL EMC ISILON F800 AND H600 I/O PERFORMANCE

DELL EMC ISILON F800 AND H600 I/O PERFORMANCE DELL EMC ISILON F800 AND H600 I/O PERFORMANCE ABSTRACT This white paper provides F800 and H600 performance data. It is intended for performance-minded administrators of large compute clusters that access

More information

Accelerating Spectrum Scale with a Intelligent IO Manager

Accelerating Spectrum Scale with a Intelligent IO Manager Accelerating Spectrum Scale with a Intelligent IO Manager Ray Coetzee Pre-Sales Architect Seagate Systems Group, HPC 2017 Seagate, Inc. All Rights Reserved. 1 ClusterStor: Lustre, Spectrum Scale and Object

More information

Introduction to High Performance Parallel I/O

Introduction to High Performance Parallel I/O Introduction to High Performance Parallel I/O Richard Gerber Deputy Group Lead NERSC User Services August 30, 2013-1- Some slides from Katie Antypas I/O Needs Getting Bigger All the Time I/O needs growing

More information

An Exploration into Object Storage for Exascale Supercomputers. Raghu Chandrasekar

An Exploration into Object Storage for Exascale Supercomputers. Raghu Chandrasekar An Exploration into Object Storage for Exascale Supercomputers Raghu Chandrasekar Agenda Introduction Trends and Challenges Design and Implementation of SAROJA Preliminary evaluations Summary and Conclusion

More information

Cloud Computing with FPGA-based NVMe SSDs

Cloud Computing with FPGA-based NVMe SSDs Cloud Computing with FPGA-based NVMe SSDs Bharadwaj Pudipeddi, CTO NVXL Santa Clara, CA 1 Choice of NVMe Controllers ASIC NVMe: Fully off-loaded, consistent performance, M.2 or U.2 form factor ASIC OpenChannel:

More information

Illinois Proposal Considerations Greg Bauer

Illinois Proposal Considerations Greg Bauer - 2016 Greg Bauer Support model Blue Waters provides traditional Partner Consulting as part of its User Services. Standard service requests for assistance with porting, debugging, allocation issues, and

More information

Evaluating New Communication Models in the Nek5000 Code for Exascale

Evaluating New Communication Models in the Nek5000 Code for Exascale Evaluating New Communication Models in the Nek5000 Code for Exascale Ilya Ivanov (KTH), Rui Machado (Fraunhofer), Mirko Rahn (Fraunhofer), Dana Akhmetova (KTH), Erwin Laure (KTH), Jing Gong (KTH), Philipp

More information

Fast Forward I/O & Storage

Fast Forward I/O & Storage Fast Forward I/O & Storage Eric Barton Lead Architect 1 Department of Energy - Fast Forward Challenge FastForward RFP provided US Government funding for exascale research and development Sponsored by 7

More information

Emerging Technologies for HPC Storage

Emerging Technologies for HPC Storage Emerging Technologies for HPC Storage Dr. Wolfgang Mertz CTO EMEA Unstructured Data Solutions June 2018 The very definition of HPC is expanding Blazing Fast Speed Accessibility and flexibility 2 Traditional

More information

Short Talk: System abstractions to facilitate data movement in supercomputers with deep memory and interconnect hierarchy

Short Talk: System abstractions to facilitate data movement in supercomputers with deep memory and interconnect hierarchy Short Talk: System abstractions to facilitate data movement in supercomputers with deep memory and interconnect hierarchy François Tessier, Venkatram Vishwanath Argonne National Laboratory, USA July 19,

More information

Multiprocessors and Thread Level Parallelism Chapter 4, Appendix H CS448. The Greed for Speed

Multiprocessors and Thread Level Parallelism Chapter 4, Appendix H CS448. The Greed for Speed Multiprocessors and Thread Level Parallelism Chapter 4, Appendix H CS448 1 The Greed for Speed Two general approaches to making computers faster Faster uniprocessor All the techniques we ve been looking

More information

Mass-Storage Structure

Mass-Storage Structure CS 4410 Operating Systems Mass-Storage Structure Summer 2011 Cornell University 1 Today How is data saved in the hard disk? Magnetic disk Disk speed parameters Disk Scheduling RAID Structure 2 Secondary

More information

Towards Automatic Heterogeneous Computing Performance Analysis. Carl Pearson Adviser: Wen-Mei Hwu

Towards Automatic Heterogeneous Computing Performance Analysis. Carl Pearson Adviser: Wen-Mei Hwu Towards Automatic Heterogeneous Computing Performance Analysis Carl Pearson pearson@illinois.edu Adviser: Wen-Mei Hwu 2018 03 30 1 Outline High Performance Computing Challenges Vision CUDA Allocation and

More information

Computer Science Section. Computational and Information Systems Laboratory National Center for Atmospheric Research

Computer Science Section. Computational and Information Systems Laboratory National Center for Atmospheric Research Computer Science Section Computational and Information Systems Laboratory National Center for Atmospheric Research My work in the context of TDD/CSS/ReSET Polynya new research computing environment Polynya

More information

Scalability issues : HPC Applications & Performance Tools

Scalability issues : HPC Applications & Performance Tools High Performance Computing Systems and Technology Group Scalability issues : HPC Applications & Performance Tools Chiranjib Sur HPC @ India Systems and Technology Lab chiranjib.sur@in.ibm.com Top 500 :

More information

GPU-centric communication for improved efficiency

GPU-centric communication for improved efficiency GPU-centric communication for improved efficiency Benjamin Klenk *, Lena Oden, Holger Fröning * * Heidelberg University, Germany Fraunhofer Institute for Industrial Mathematics, Germany GPCDP Workshop

More information

SSD/Flash for Modern Databases. Peter Zaitsev, CEO, Percona November 1, 2014 Highload Moscow,Russia

SSD/Flash for Modern Databases. Peter Zaitsev, CEO, Percona November 1, 2014 Highload Moscow,Russia SSD/Flash for Modern Databases Peter Zaitsev, CEO, Percona November 1, 2014 Highload++ 2014 Moscow,Russia Percona We love Open Source Software Percona Server Percona Xtrabackup Percona XtraDB Cluster Percona

More information

CS500 SMARTER CLUSTER SUPERCOMPUTERS

CS500 SMARTER CLUSTER SUPERCOMPUTERS CS500 SMARTER CLUSTER SUPERCOMPUTERS OVERVIEW Extending the boundaries of what you can achieve takes reliable computing tools matched to your workloads. That s why we tailor the Cray CS500 cluster supercomputer

More information

S8765 Performance Optimization for Deep- Learning on the Latest POWER Systems

S8765 Performance Optimization for Deep- Learning on the Latest POWER Systems S8765 Performance Optimization for Deep- Learning on the Latest POWER Systems Khoa Huynh Senior Technical Staff Member (STSM), IBM Jonathan Samn Software Engineer, IBM Evolving from compute systems to

More information

Functional Partitioning to Optimize End-to-End Performance on Many-core Architectures

Functional Partitioning to Optimize End-to-End Performance on Many-core Architectures Functional Partitioning to Optimize End-to-End Performance on Many-core Architectures Min Li, Sudharshan S. Vazhkudai, Ali R. Butt, Fei Meng, Xiaosong Ma, Youngjae Kim,Christian Engelmann, and Galen Shipman

More information

NVMe Takes It All, SCSI Has To Fall. Brave New Storage World. Lugano April Alexander Ruebensaal

NVMe Takes It All, SCSI Has To Fall. Brave New Storage World. Lugano April Alexander Ruebensaal Lugano April 2018 NVMe Takes It All, SCSI Has To Fall freely adapted from ABBA Brave New Storage World Alexander Ruebensaal 1 Design, Implementation, Support & Operating of optimized IT Infrastructures

More information

DDN. DDN Updates. DataDirect Neworks Japan, Inc Nobu Hashizume. DDN Storage 2018 DDN Storage 1

DDN. DDN Updates. DataDirect Neworks Japan, Inc Nobu Hashizume. DDN Storage 2018 DDN Storage 1 1 DDN DDN Updates DataDirect Neworks Japan, Inc Nobu Hashizume DDN Storage 2018 DDN Storage 1 2 DDN A Broad Range of Technologies to Best Address Your Needs Your Use Cases Research Big Data Enterprise

More information

RAMCloud and the Low- Latency Datacenter. John Ousterhout Stanford University

RAMCloud and the Low- Latency Datacenter. John Ousterhout Stanford University RAMCloud and the Low- Latency Datacenter John Ousterhout Stanford University Most important driver for innovation in computer systems: Rise of the datacenter Phase 1: large scale Phase 2: low latency Introduction

More information

Multi-Threaded UPC Runtime for GPU to GPU communication over InfiniBand

Multi-Threaded UPC Runtime for GPU to GPU communication over InfiniBand Multi-Threaded UPC Runtime for GPU to GPU communication over InfiniBand Miao Luo, Hao Wang, & D. K. Panda Network- Based Compu2ng Laboratory Department of Computer Science and Engineering The Ohio State

More information

File system internals Tanenbaum, Chapter 4. COMP3231 Operating Systems

File system internals Tanenbaum, Chapter 4. COMP3231 Operating Systems File system internals Tanenbaum, Chapter 4 COMP3231 Operating Systems Summary of the FS abstraction User's view Hierarchical structure Arbitrarily-sized files Symbolic file names Contiguous address space

More information

CSCI-GA Database Systems Lecture 8: Physical Schema: Storage

CSCI-GA Database Systems Lecture 8: Physical Schema: Storage CSCI-GA.2433-001 Database Systems Lecture 8: Physical Schema: Storage Mohamed Zahran (aka Z) mzahran@cs.nyu.edu http://www.mzahran.com View 1 View 2 View 3 Conceptual Schema Physical Schema 1. Create a

More information

Do You Know What Your I/O Is Doing? (and how to fix it?) William Gropp

Do You Know What Your I/O Is Doing? (and how to fix it?) William Gropp Do You Know What Your I/O Is Doing? (and how to fix it?) William Gropp www.cs.illinois.edu/~wgropp Messages Current I/O performance is often appallingly poor Even relative to what current systems can achieve

More information

RAIN: Reinvention of RAID for the World of NVMe

RAIN: Reinvention of RAID for the World of NVMe RAIN: Reinvention of RAID for the World of NVMe Dmitrii Smirnov Principal Software Developer smirnov.d@raidix.com RAIDIX LLC 1 About the company RAIDIX is an innovative solution provider and developer

More information

Oracle Performance on M5000 with F20 Flash Cache. Benchmark Report September 2011

Oracle Performance on M5000 with F20 Flash Cache. Benchmark Report September 2011 Oracle Performance on M5000 with F20 Flash Cache Benchmark Report September 2011 Contents 1 About Benchware 2 Flash Cache Technology 3 Storage Performance Tests 4 Conclusion copyright 2011 by benchware.ch

More information

RAIN: Reinvention of RAID for the World of NVMe. Sergey Platonov RAIDIX

RAIN: Reinvention of RAID for the World of NVMe. Sergey Platonov RAIDIX RAIN: Reinvention of RAID for the World of NVMe Sergey Platonov RAIDIX 1 NVMe Market Overview > 15 vendors develop NVMe-compliant servers and appliances > 50% of servers will have NVMe slots by 2020 Market

More information

The Cray Rainier System: Integrated Scalar/Vector Computing

The Cray Rainier System: Integrated Scalar/Vector Computing THE SUPERCOMPUTER COMPANY The Cray Rainier System: Integrated Scalar/Vector Computing Per Nyberg 11 th ECMWF Workshop on HPC in Meteorology Topics Current Product Overview Cray Technology Strengths Rainier

More information

UK LUG 10 th July Lustre at Exascale. Eric Barton. CTO Whamcloud, Inc Whamcloud, Inc.

UK LUG 10 th July Lustre at Exascale. Eric Barton. CTO Whamcloud, Inc Whamcloud, Inc. UK LUG 10 th July 2012 Lustre at Exascale Eric Barton CTO Whamcloud, Inc. eeb@whamcloud.com Agenda Exascale I/O requirements Exascale I/O model 3 Lustre at Exascale - UK LUG 10th July 2012 Exascale I/O

More information

I/O Profiling Towards the Exascale

I/O Profiling Towards the Exascale I/O Profiling Towards the Exascale holger.brunst@tu-dresden.de ZIH, Technische Universität Dresden NEXTGenIO & SAGE: Working towards Exascale I/O Barcelona, NEXTGenIO facts Project Research & Innovation

More information

Advanced Software for the Supercomputer PRIMEHPC FX10. Copyright 2011 FUJITSU LIMITED

Advanced Software for the Supercomputer PRIMEHPC FX10. Copyright 2011 FUJITSU LIMITED Advanced Software for the Supercomputer PRIMEHPC FX10 System Configuration of PRIMEHPC FX10 nodes Login Compilation Job submission 6D mesh/torus Interconnect Local file system (Temporary area occupied

More information

DDN s Vision for the Future of Lustre LUG2015 Robert Triendl

DDN s Vision for the Future of Lustre LUG2015 Robert Triendl DDN s Vision for the Future of Lustre LUG2015 Robert Triendl 3 Topics 1. The Changing Markets for Lustre 2. A Vision for Lustre that isn t Exascale 3. Building Lustre for the Future 4. Peak vs. Operational

More information

BeeGFS Solid, fast and made in Europe

BeeGFS Solid, fast and made in Europe David Ramírez Alvarez HPC INTEGRATOR MANAGER WWW.SIE.ES dramirez@sie.es ADMINTECH 2016 BeeGFS Solid, fast and made in Europe www.beegfs.com Thanks to Sven for info!!!, CEO, ThinkParQ What is BeeGFS? BeeGFS

More information

Virtual Memory. Reading. Sections 5.4, 5.5, 5.6, 5.8, 5.10 (2) Lecture notes from MKP and S. Yalamanchili

Virtual Memory. Reading. Sections 5.4, 5.5, 5.6, 5.8, 5.10 (2) Lecture notes from MKP and S. Yalamanchili Virtual Memory Lecture notes from MKP and S. Yalamanchili Sections 5.4, 5.5, 5.6, 5.8, 5.10 Reading (2) 1 The Memory Hierarchy ALU registers Cache Memory Memory Memory Managed by the compiler Memory Managed

More information

Moneta: A High-Performance Storage Architecture for Next-generation, Non-volatile Memories

Moneta: A High-Performance Storage Architecture for Next-generation, Non-volatile Memories Moneta: A High-Performance Storage Architecture for Next-generation, Non-volatile Memories Adrian M. Caulfield Arup De, Joel Coburn, Todor I. Mollov, Rajesh K. Gupta, Steven Swanson Non-Volatile Systems

More information

CSE 421/521 Final Exam

CSE 421/521 Final Exam Name UBID Seat Question: 1 2 3 4 5 6 7 8 9 10 Total Points: 10 5 5 5 5 5 5 20 25 25 100 Score: CSE 421/521 Final Exam 09 May 2016 Please fill out your name and UB ID number above. Also write your UB ID

More information

Block Device Scheduling. Don Porter CSE 506

Block Device Scheduling. Don Porter CSE 506 Block Device Scheduling Don Porter CSE 506 Logical Diagram Binary Formats Memory Allocators System Calls Threads User Kernel RCU File System Networking Sync Memory Management Device Drivers CPU Scheduler

More information

Block Device Scheduling

Block Device Scheduling Logical Diagram Block Device Scheduling Don Porter CSE 506 Binary Formats RCU Memory Management File System Memory Allocators System Calls Device Drivers Interrupts Net Networking Threads Sync User Kernel

More information

Chapter 10: Mass-Storage Systems

Chapter 10: Mass-Storage Systems COP 4610: Introduction to Operating Systems (Spring 2016) Chapter 10: Mass-Storage Systems Zhi Wang Florida State University Content Overview of Mass Storage Structure Disk Structure Disk Scheduling Disk

More information

Chapter 11: File-System Interface

Chapter 11: File-System Interface Chapter 11: File-System Interface Silberschatz, Galvin and Gagne 2013 Chapter 11: File-System Interface File Concept Access Methods Disk and Directory Structure File-System Mounting File Sharing Protection

More information

Week 12: File System Implementation

Week 12: File System Implementation Week 12: File System Implementation Sherif Khattab http://www.cs.pitt.edu/~skhattab/cs1550 (slides are from Silberschatz, Galvin and Gagne 2013) Outline File-System Structure File-System Implementation

More information

Red Hat Enterprise 7 Beta File Systems

Red Hat Enterprise 7 Beta File Systems Red Hat Enterprise 7 Beta File Systems New Scale, Speed & Features Ric Wheeler Director Red Hat Kernel File & Storage Team Red Hat Storage Engineering Agenda Red Hat Enterprise Linux 7 Storage Features

More information

Next Generation Architecture for NVM Express SSD

Next Generation Architecture for NVM Express SSD Next Generation Architecture for NVM Express SSD Dan Mahoney CEO Fastor Systems Copyright 2014, PCI-SIG, All Rights Reserved 1 NVMExpress Key Characteristics Highest performance, lowest latency SSD interface

More information

Mass-Storage Structure

Mass-Storage Structure Operating Systems (Fall/Winter 2018) Mass-Storage Structure Yajin Zhou (http://yajin.org) Zhejiang University Acknowledgement: some pages are based on the slides from Zhi Wang(fsu). Review On-disk structure

More information

Lecture 17: Threads and Scheduling. Thursday, 05 Nov 2009

Lecture 17: Threads and Scheduling. Thursday, 05 Nov 2009 CS211: Programming and Operating Systems Lecture 17: Threads and Scheduling Thursday, 05 Nov 2009 CS211 Lecture 17: Threads and Scheduling 1/22 Today 1 Introduction to threads Advantages of threads 2 User

More information

Store Process Analyze Collaborate Archive Cloud The HPC Storage Leader Invent Discover Compete

Store Process Analyze Collaborate Archive Cloud The HPC Storage Leader Invent Discover Compete Store Process Analyze Collaborate Archive Cloud The HPC Storage Leader Invent Discover Compete 1 DDN Who We Are 2 We Design, Deploy and Optimize Storage Systems Which Solve HPC, Big Data and Cloud Business

More information

Lustre overview and roadmap to Exascale computing

Lustre overview and roadmap to Exascale computing HPC Advisory Council China Workshop Jinan China, October 26th 2011 Lustre overview and roadmap to Exascale computing Liang Zhen Whamcloud, Inc liang@whamcloud.com Agenda Lustre technology overview Lustre

More information

OPERATING SYSTEM. Chapter 12: File System Implementation

OPERATING SYSTEM. Chapter 12: File System Implementation OPERATING SYSTEM Chapter 12: File System Implementation Chapter 12: File System Implementation File-System Structure File-System Implementation Directory Implementation Allocation Methods Free-Space Management

More information

Fusion Engine Next generation storage engine for Flash- SSD and 3D XPoint storage system

Fusion Engine Next generation storage engine for Flash- SSD and 3D XPoint storage system Fusion Engine Next generation storage engine for Flash- SSD and 3D XPoint storage system Fei Liu, Sheng Qiu, Jianjian Huo, Shu Li Alibaba Group Santa Clara, CA 1 Software overhead become critical Legacy

More information

IBM Deep Learning Solutions

IBM Deep Learning Solutions IBM Deep Learning Solutions Reference Architecture for Deep Learning on POWER8, P100, and NVLink October, 2016 How do you teach a computer to Perceive? 2 Deep Learning: teaching Siri to recognize a bicycle

More information

TECHNICAL GUIDELINES FOR APPLICANTS TO PRACE 6 th CALL (Tier-0)

TECHNICAL GUIDELINES FOR APPLICANTS TO PRACE 6 th CALL (Tier-0) TECHNICAL GUIDELINES FOR APPLICANTS TO PRACE 6 th CALL (Tier-0) Contributing sites and the corresponding computer systems for this call are: GCS@Jülich, Germany IBM Blue Gene/Q GENCI@CEA, France Bull Bullx

More information