Download to read offline
Uploaded byTyrone Systems
Introduction to HPC & Supercomputing in AI
The document outlines a webinar on High-Performance Computing (HPC) hosted by Netweb, detailing their systems, services, and solutions for HPC workloads, including GPU optimized AI/ML solutions and unified storage options. It introduces the Tyrone Cluster Manager (TCM) that facilitates management of HPC clusters and discusses various applications supported by their hardware. Key services include scalable storage solutions, deep learning capabilities, and customization through an online product configurator.
More Related Content
Similar to Introduction to HPC & Supercomputing in AI
Introduction to HPC & Supercomputing in AI
- 1. Modern way tolook at HPC Workloads Monday | 25TH MAY, 2020 LIVE WEBINAR Presented by
- 2. Agenda • About Netweb •HPC o TCM o Storage • Systems Offerings • Interconnect • GPU Optimized AI | ML Solutions • Converged Mixed Workloads (Future of HPC) • Tyrone Kubits™ • Q&A Session
- 3. Tyrone Systems ata Glance
- 4. A State ofthe Art Facility at Netweb HQ Tyrone -100Gbps Interconnect based R&D Cluster with PFS HPC Experts part of Netweb Team In House Supercomputing Lab 300+ HPC Nodes Test Facility HPC observatory LABHPC Burn in Test Lab Netweb HQ@ Faridabad
- 5. Solutions that spanthe entire Data Center SERVER • HPC Servers • Mission Critical X86 • Storage Servers • High-Density Servers • GPU Servers Cloud Solutions Big Data/AIHPC Solutions Cloud Big Data Virtualization AI / DEEP LEARNING Product Portfolio WORKSTATIONS • GPU Workstations • Tower | Rack • Liquid Cooling STORAGE • Unified Storage • Storage Array • Archival • JBOD • Ceph Storage NETWORKING • InfiniBand • Omnipath Architecture Tyrone Kubernetes Platform HPC Cluster GPU Optimised Supercomputer HPC On Cloud SMP Solutions Mngmt Tools Analytics Data Insights HPC cluster parallel file systems Inferencing Hyper-converged Virtual SAN Mixed Workloads GPU Systems
- 6.
- 7. HPC Architecture Master Node-1Master Node-2 Parallel File System I/O I/O Storage IB/OPA Switch Control Node 10GbE Switch Node - 1 Node - 2 Node - 3 Node - 4
- 8. Our HPC Services TheParallel Application Applications End Users, ISV’s Job Control Batch queue, schedulers, cluster monitoring, cluster control Communication Libraries (PVM, MPI, etc)Middleware Open Source / Paid HPC Cluster Community, Interconnect Networking Components Compute Node The Operating system (Linux etc)OS OSV’s Master Node Proposing hardware Designing HPC architecture Application level support Post Cluster Installation Services Inter-node communications hardware/software Processor base, physical format, H/W management, etc. Interfacing b/w the entire cluster and user environment
- 9.
- 10.
- 11. TCM- Architectural Overview ASingle Dashboard has a representation of: • The number of Nodes along with all node names, Groups, Users and User-Groups. • Configuration Processes like Ganglia, torque, Slurm and PBS. • The Graphs represents CPU Utilization, Disk Free, Memory Shared and Network Details. • List of all the processes which we have pushed in the backend for processing through celery and their results.
- 12. TCM- Configuration Processes Differentmodules that can be configured with the help of TCM. The modules are • Create TCM repo • Ganglia • Torque • Slurm • PBS • Installing Driver(Mellanox and OPA) • Shared Home – AutoFS We can install their master as well as client and we can also Uninstall as per our specification.
- 13. TCM- GCC Applications Thelater version of TCM has the option to compile GCC applications with the help of our WEB Interface. • BGW • Gromacs • NCView • Transiesta • Cp2k • Hdf5 • Openbabel • Openblas • Yambo • Cpmd • FFTW • Openmpi • Grads • Namd • Siesta • Abinit • Hdf5 • ShengBTE • BGW • Mvapich2 • Siesta • Cp2k • Nwchem • TranSiesta • Gromacs • QE 62 The application which can be installed in GCC are as follows:
- 14.
- 15. HIGH PERFORMANCE PARALLELFILE STORAGE • Supports Large Datasets & High IOPS Requirements • Large, fast distributed scratch file system • A centralised storage for clusters • Simple building block architecture delivers predictable scaling to specific requirements. • Industry leading storage density with High Availability design • Total Solution including Systems, Software and Services • No single point of failure • Reduce storage costs by up to 90% KEY FEATURES 10GB/s EDR: 100Gb/s HDR: 200Gb/s
- 16. SAN NAS VTL UnifiedStorage Solution : Opslag FS2 Key Features 4GB/s Bandwidth 576 TB in Single 4U Enclosure 5+PB Scalability Native InfiniBand Support (SRP/NFS over RDMA) All-in-one-solution (NAS/FC/iSCSI/SRP/VTL) File Access Protocol CIFS/SMB AFP FTP NFS NFS over RDMA Block Access Protocol iSCSI Target FC Target SRP Target
- 17. Key Benefits Scalability :Easily scale from a few terabytes to 100+ petabytes Unified Storage: SAN/NAS & VTL combines file, block & Object storage into a single storage box Unified Management: Single interface for centralized management Unified Protection: Single solution for local and remote application protection Dual Controller provides an extra level of data protection Connects to Any network : Gigabit Ethernet, Fibre-Channel, FDR/EDR InfiniBand Intel Omni- path Supports SSD caching for customers looking for extremely high IOPS. Our SSD caching Deduplication & Compression manage actual required capacity & cost HIGH Capacity HIGH Performance FS2 can be set up as All-Flash or Hybrid Flash storage
- 18. Our System Offerings StorageGPU 4U8Node 2U 4NodeAll-Flash NVme
- 19.
- 20.
- 21. PFS Design- ArchitecturalOverview • 19’’ rack mountable 1U chassis with Dual redundant slots • 40 QSFP56 non-blocking ports with aggregate data throughput up to 16 Tb/s (HDR) • Management Ports – 100/1000 RJ45 Ethernet port • Connectors & Cabling • QSFP56 connectors • Passive copper or active fiber cables
- 22. GPU optimized forAI/ML Solutions
- 23. Delivers 4XFASTER TRAINING thanother GPU-based systems Your Personal AI Supercomputer Power-on to Deep Learning in Minutes Pre-installed with Powerful Deep Learning Software Extend workloads from your Desk-to-Cloud in Minutes
- 24. GPU Systems OptimizedFor Deep Learning GPUS 1 X GPU 2 X GPUs 3 x GPUs 4 x GPUs 6 x GPUs 8 x GPUs 10 x GPUs 16 x GPUs 20 x GPUs MODEL SS400TR-54R SS400TG-16T DS400TG-14R DS400TG-48R DS400TG-12RT DS400TG-12RT DS400TGH-28R DS400TQV- 416RT DS400TOG- 424R DS400TOG- 424RT DS400NG16- 1016RT DS400TG- 424RT FORM FACTOR 5U 1U 1U 4U 1U 1U 1U 4U 4U 4U 10U 4U COMPUTE PERFORMANCE 1 X Tesla V100 32 Single Precision 14+ TFs 2 X Tesla V100 32 Single Precision 28+ TFs 3 X Tesla V100 32 Single Precision 42+ TFs 4 X Tesla V100 32 Single Precision 56+ TFs 4 X Tesla V100 32 Single Precision 56+ TFs 4 X Tesla V100 32 Single Precision 56+ TFs 6 X Tesla V100 32 Single Precision 84+ TFs 8 X Tesla V100 32 Single Precision 125+ TFs 8 X 2080 Ti Single Precision 100+ TFs 8 X Tesla V100 32 Single Precision 100+ TFs 10 X 2080 Ti Single Precision 130+ TFs 10 X Tesla V100 32 Single Precision 140+ TFs 16 X Tesla V100 32 Single Precision 250+ TFs 20 X T4 GPUs Single Precision 160+ TFs FP16/FP32 Mixed Precision 1300+ TFs TYRONE KUBITS ACCESS NUMBER OF GPU’S COMPUTEPERFORMANCE
- 25. FASTER AI INNOVATION& INSIGHT
- 26. Mixed Workloads Convergenceof AI|HPC| Cloud | Containers
- 27. The Era ofMixed Workload F L E X I B L E Is the usage going to be constant? O P T I M I Z A T I O N Is optimal utilization required? R E S I L I E N C E Do we need the application to run all the time. E A S E Is ‘ease of maintenance’ key? S C A L A B I L I T Y & S P E E D Do we have one size that fits all?
- 28. Connectivity and usage VirtualDesktop Laptop Tyrone Cloud Manager Tyrone Cloud Manager Laptop
- 29. Expand Cloud Tyrone Hardware OS TyroneCloud Suite (TCS) Service management Log management Monitoring OPENSTACK SHARED SERVICES KeyStone, Horizon, Ceilometer, Percona XtraDB, RabbitMQ, Memcache, MongoDB, Kubernetes KVM/LXD Nova Dockers Network -Neutron Ceph Storage Image Block Object / Swift
- 30. GPU CONTAINERIZED APPLICATION DEEP LEARNINGAPPLICATIONS DEEP LEARNING FRAMEWORKS DEEP LEARNING LIBRARIES CUDA TOOLKIT MAPPED NVIDIA DRIVERS CONTAINER OS CONTAINERIZATION TOOL DOCKER ENGINE NVIDIA DRIVER HOST OS NVIDIA CONTAINER RUNTIME FOR DOCKER SERVER INFRASTRUCTURE HOST OPERATING SYSTEM DOCKER ENGINE Bins / Libs Bins / Libs Bins / Libs App1 App1 App1
- 31. Run Multiple Applications simultaneously TyroneKUBITS™ Cloud Flow Architecture Revolutionizing Deep Learning CPU-GPU Environment KUBITS™ Compatible Workstations WITH TYRONE KUBITS™ CLIENT KUBITS has a repository of : 50 containerized applications 100s of Containers 10X20X30X40X50X60X70X SPEED
- 32. Tyrone KUBITS :Revolutionizing Deep Learning CPU-GPU Environment Run different applications simultaneously Check for Tyrone KUBITS Compatible Workstations Get access to over 100+ Containers on Tyrone KUBITS Cloud. High scalability Affordable price Has both GPU & CPU Optimized Containers Design a simple Workstation or Large Clusters with KUBITS technology. Talk to our experts & build the right workstation within your budget. KUBITS CLOUDCOMPATIBLE
- 33. Cloud Deployment in5 countries INDIA SINGAPORE UK USA AUSTRALIA
- 34. Product Configurator • Aneasy to use tool on our Tyrone Website • Allows customers to select from over 400 products and provides customized solution • Easy to view all SKUs needed to quote particular End User • Easy access to product technical specifications http://tyronesystems.com/servers/servers_workstation.html
- 35. Q&A SessionContact ourteam if you have any further questions after this webinar ai@netwebtech.comTalk to our AI Experts Navin navin@netwebindia.com Tushar tushar@netwebindia.com Anurag Anurag.thakare@netwebindia.com Anjani anjani.pandey@netwebindia.com