Using HPC Computational Physics Tools for Advanced Engineering Simulations and Production Deployment David Hinz Sr. Director, IT Engineering Services / HPC Byron Lengsfield III Technologist, Recording Subsystems March 26, 2018 ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 Disclaimer • This presentation may contain product features that are currently under development. • This overview of new technology represents no commitment from VMware to deliver these features in any generally available product. • Features are subject to change, and must not be included in contracts, purchase orders, or sales agreements of any kind. • Technical feasibility and market demand will affect final delivery. • Pricing and packaging for any new technologies or features discussed or presented have not been determined. 2 Agenda • Western Digital’s History with AWS and GPUs for HPC • Current Uses of AWS GPUs • Recommendations for Success Confidential 3 Western Digital: A Heritage of Innovation 14,000+ active patents worldwide 1956 | IBM Invents First Hard Drive (RAMAC) 2017 | SanDisk 1976 | WD First 96-layer 3D NAND Technology First Disk Array Sub-System Patent 2016 | SanDisk First 64-layer 3D NAND Technology 1988 | SunDisk Introduces Flash as a Storage Medium 2015 | SanDisk First 48-layer 3D NAND Technology 1991 | IBM 2013 | HGST Introduces MR Heads First Helium HDD (6TB) 1995| M-Systems 2013 | HGST First Flash Drive First 12Gb/s SAS SSD 2009 | SanDisk 2000| IBM & M-Systems First 64GB X4 Flash memory First USB Flash Drive 2009 | Fusion-io 2000 | SanDisk First Enterprise PCIe Flash Accelerators First MLC NAND 2006| HGST 2008| G-Technology First 2.5-inch HDD First 1TB Portable PMR Technology External HDD ©2018 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 4 Portfolio of Leading Brands ©2018 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 5 Leading Portfolio Breadth and Depth Solutions to Capture, Preserve, Access and Transform Data Data Center Client Devices Client Solutions Devices & Solutions Platforms/Systems/ Solutions Devices Media Device RISC-V Heads System Technologies Software Controllers Software ©2018 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 6 Fueling the Next Decade of Big Data New technology will enable 40TB by 2025 and even higher beyond ©2018 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 7 Western Digital Cloud HPC Architecture • Dedicated AWS account for HPC – DirectConnect = Main – VPN = Backup • Clusters for the usage queue – By Organizations Cluster – By Applications SSH & VNC Master Computing Client data I/O Nodes – By Users Users attached • Autoscaling Cluster WDC – Autoscale based on the user queue NFS Shared – CycleCloud by CycleComputing Cycle Storage – Master node = Job Scheduler + NFS Cloud – Instance Store for NFS = i3 instance Oregon – Optimal Instance Type for the workload = c4, r4, p2 ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 8 Shape Compute To Match Work To Be Done Molecular Dynamics 1.67x Initial Overall Throughput Gain Fluid Dynamics 1.4x Overall Throughput Gain Molecular Dynamics MicroMagnetics Throughput Parameter Sweeps Throughput Gain Simulation Type Increase Model 1 1.23x - 1.78x Head Drive Interface Vacuum 1.99x Gaps Model 2 1.01x – 1.67x Vacuum Gap "collection" 4.00x Model 3 1.23x – 1.69x Media Grains for HAMR (FePt/C) 2.03x Model 4 Up to 2.7x 4 Carbon Molecule Clusters 5.67x ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 9 Western Digital’s Innovations + the Cloud Advantage = Faster Business Value HDD Arial Density Mechanical INNOVATIONS 6TB Number of Disks Cloud Air/Helium CLOUD Speed ADVANTAGE 4TB Lubes HPC Key to Scalability Enabling and Driving 85,000+ Cores Innovation 3+ Data Centers High Performance HPC CPU HPC GPU +6,000 Servers Computing (HPC) Cluster Cluster On-premise HGST runs worlds largest F500 cloud HPC simulation run Reduces simulation time from 30 days to 8 hours!!! Confidential 10 Western Digital : AWS and GPUs for HPC • Western Digital has been a partner with AWS to understand workloads and use cases of HPC GPU computing for 3+ years • Western Digital tested AWS GPU products and provided detailed performance benchmarks • Western Digital directly engaged GPU suppliers directly to highlight use cases and requirements • Actively using AWS GPUs for Machine Learning / Artificial Intelligence and High Performance Computing Confidential 11 HDD Recording System Modeling: Outline n Modeling Work Flow n GPU/AWS Realization of the Work Flow n Model Details v Parallelization yields large gains vLarger problems: 511 bit patterns, which allows to use a common set of signal processing tools in experiment and in the modeling vOptimization in a large parameter space vDescribe recording physics at a smaller length scale n Conclusions v Seamless migration to AWS v Realization of the capabilities of a new generation of GPUs ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 12 Hard Disk Drive (HDD) Magnetic Recording … 0 0 0 1 1 1 1 0 1 1 1 1 0 1 0 0 0 … 1. Write step Write Head Read Head 2. Read step 3. Signal Processing Step Head Motion Re Read Sensor 1 1 1 1 1 1 1 1 0 1 … 0 0 0 0 0 0 0 … § Input: 2-level write current § Output: Noisy readback waveform Tracks of data written § Goal: Reliably store data at highest on the disk possible areal density ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 13 Work Flow Current waveforms input to head Write Current Waveforms model to produce write field in the media as a function of time (1500, 3D head fields) Multilayer micromagnetic media models used to generate 400 repeats of Pseudo-Random Bit Pattern (PRBS) which is then Writer Model written by the write field, read and fed into a SWC for BER and other metrics Micromagnetic or FEM Writer Model for Head Fields Write Field Micromagnetic Multi-layer Granular Magnetic Media Write Process: Multi-Layer Media Model (GPUs) - 3D Media Model - Many Coupled Layers - Thermal Field Reader Software Channel (SWC) Sensitivity Readback Function Readback SIgnal Bit Error Rate (BER) and Analysis: Signal-to-Noise (SNR), Areal Density Capability (ADC), Distortions ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 14 Fields generated from FEM or Micromagnetic Writer Model Head Motion Many critical dimensions and angles Wrap-Around-Shield (WAS) Write Head-Under-Layer Pole Spacing (HUS) trailing portion of WAS Gap between write pole and trailing portion of WAS Soft Under Layer (SUL) Write field magnitude and angle varies Competing magnetic flux paths, in all three dimensions flux can flow to the SUL or the shields Magnetic field produced in each media layer FEM = finite-element-method , magnetostatic model ©2017 Western Digital Corporation or itsFEM affiliates. All =rights finite reserved. Confidential.-element-method , magnetostatic model 3/26/18 15 Readback Process Reader Senstivity Function (RSF) 1 100 Cross Track rd.rwReadG22W26.6HUS36.6tSUL50x550dx2.5mu50.PRT 0.9 1.2 3D Reader Sensitivity Function (RSF): 0.8 1 50 0.7 Convolution of the RSF and magnetization 0.6 0.8 0 0.5 0.6 0.4 0.4 -50 0.3 1. Analytic or micromagnetic program to generate RSF 0.2 0.2 0.1 2. Readback is very fast -100 0 Ø All crosstrack positions are obtained at one time 0 MRW= 26.9923 -50 0 50 100 -0.2 using 2D FFT -300 -200 -100 0 100 200 300 Ø Post-processing step, all magnetization data is on a separate server. Ø Convolve 3D RSF with the magnetization. 3. Full micromagnetic readback is available but more compute intensive Different RSF for each media layer 2D Cut of the RSF Signal RSF Magnetization ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 16 Multi-layer Media Model: Exchange between layers and grains Top surface of the media Magnetization aligned perpendicular to the surface of the media along anisotropy axis |k> . 2 Ea = K <m|k> Macro-spin Model ( each layer is a macro-spin ) Ø Long range magnetic field computed using FFT CUDA library Ø Exchange interaction is parallelized over 3D regions ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 17 Batch System Job Flow Linux Scripts Generate Input from Templates which are Submitted to the Batch System • The input template contains hardwired input values and variables which are specified in the Submit Script. Nvidia GPUs: Coarse Grain Parallel CPU GPU Master Node Master Node Post-Processing Node Submit scripts Processing step generate an The batch system distributes (readback) to ensemble of inputs Output is moved to the command files on the generate BER, SNR, and on the master compute nodes.. the master node ADC,… node AWS ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 18 GPU code implementation • Example: Anisotropy field ( one body: <M|K> GPU|K> field code term )implementation üGrain-wise anisotropy fields computation are assigned to different cores on a GPU üEach layer in a grain runs independently • CUDA Random Numbers for Thermal Field • CUDA FFT for Magnetic Field • Parallel 3D Group Exchange GPU Media ©2017 Western Digital Corporation or its affiliates. All rights reserved. Confidential. 3/26/18 19 Analysis of Small Bit Pattern ~ 620