Get the Best out of NVIDIA GPUs for 3D Design and Engineering in the Cloud Andrea.Rodolico@nice-software.com CTO & Co-founder S5415
About NICE o o o Company Focus on technical computing since 1996 Partners all around the world Always profitable and self-funded Expertise Industry veterans in HPC Private and Public Cloud solutions Core business: Access to Grid / HPC / Cloud solutions Work, visualize, and collaborate in HPC EnginFrame and DCV product families 2
Technical Applications in the Cloud WS PC Collaboration HPC 3D Visualization Linux Job submission 3
4 The Virtual Workstation User Experience
5 Our partners
About NICE Desktop Cloud Visualization Engineers and Designers Thin viewer Dynamic Quality Controls Transfer Pixels, not Data! DCV protocol over LAN or WAN Virtual Workstation Server(s) Storage 3D Apps Remote desktop streaming protocol with optimized 3D management, leveraging NVENC H.264 Direct X and OpenGL support Broad OS and device support Server: Windows 7, 2008, RHEL, SUSE, CentOS Client: Windows, Linux, OS X, ios, Android, Web client Dynamic Bandwidth optimization balancing quality Vs. frame rate, with lossless quality updates Encrypted communication Real time collaboration with independent bandwidth control Multi-monitor support, tested for 4K Remote USB support, Smartcard support Supports multiple HyperVisors, including GPU pass-through and vgpu solutions Full Linux and Windows session management via EnginFrame Views
NVENC Accelerated H.264 Results DCV is the first technology to leverage the specialized H.264 hardware encoder on NVIDIA Kepler cards! Codec JPEG H264 BW ~5 MB/s ~1 MB/s CPU load ~16% ~4% Test setup: 1600x1200, 50% quality, 25 fps cap, G2 instance on AWS H.264 powers both Linux and Windows remote graphics!
Our architecture Access Thin viewer Self-Service Offering Resources Collaborators, Support staff Remote 3D desktop protocol Visualization Servers Linux & Windows 3D sessions HTTP(S) End Users Command-line Storage SOAP HPC schdulers HPC jobs Developers, Integrators
Pure-Virtualization Architecture Clients can run Linux, Windows or Mac OS/X Standard TCP/IP network Dynamic compression Encrypted communication DCV DCV DCV vgpu Windows VM-1 GPU pass-through Linux VM-2 Hypervisor with vgpu support Server equipped with NVIDIA GRID K1/K2 vgpu Windows VM-n Application in the VM run with native NVIDIA driver GPU memory is split between users All other resources are shared GRID K2 recommended for CAD / CAE / Technical Viz.
HPC-Optimized Architecture Clients can run Linux, Windows or Mac OS/X Standard TCP/IP network Dynamic compression Encrypted communication DCV DCV DCV DCV Library Linux VDesktop-1 DCV Library Linux VDesktop-2 RHEL / CentOS / SuSE with DCV Rendering Server System equipped with NVIDIA Quadro/Tesla DCV Library Windows KVM-n Windows application in the VM using DCV driver Linux applications running in Virtual Desktops All resources are shared HPC schedulers support Kepler cards recommended for GPGPU performance
11 Technical Computing in the Public Cloud
ANSYS: Regional HPC and Pre-post processing o Problem Scattered HPC facilities Widely distributed users Large modes / mesh travelling over WAN Customer demand for centralization solutions o Solution NICE DCV multi-platform NICE EnginFrame Single visualization & HPC cluster refresh Certification of DCV for ANSYS main products o Benefits Regional data/app consolidation Minimize WAN data transfers Home working over VPN Standardized support for customer solutions Improved support quality Public Cloud foundation
Leveraging the Public Cloud for product testing o AWS currently offers G2 instances Graphics oriented, but low memory 1500+ GPU cores NVIDIA GRID powered o AWS Test Drive at a glance Visible in AWS Test Drive site: http://aws.amazon.com/testdrive/hpc/ Completely self-service for end user Leveraged DCV external rendering to overcome memory limitations 13
Large Memory Vs. GPU? Why not both? Graphics workload R3 End User DCV Proxy Thin viewer DCV protocol over HTTP(S) G2 14
Major EU Automotive: CAE consolidation o Problem Growing CAE data Large distributed user base Memory and bandwidth shortage on workstations Decreasing engineer productivity o Solution NICE DCV on Linux NICE EnginFrame High-end, large memory CAE resources Smart sharing policies o Benefits Greatly improved speed to access data No memory bottleneck Better productivity and responsiveness Thin clients, BYOD ready Savings on workstation upgrade or refresh
NORDAM: Catia on Demand o Problem As a Tier 1 OEM supplier, needs to match customers Catia revisions and SP s Difficult on a workstation o Benefits Allows me to easily upgrade, and test Service Packs OS is not tied to physical hardware, allows for greater flexibility Thin / remote client ready o Solution NICE DCV & EnginFrame XenServer w/ vgpu support NVIDIA GRID K2 16
ENI: Global E&P Cloud o Problem WW distributed operations Experts at HQ or tavelling, many users at subsidiaries Very large models can t be moved or synchronized Variable WAN latency Linux and Windows apps o Solution NICE DCV & HP RGS NICE EnginFrame Multiple hubs with 3D and HPC resources o Benefits U S A Ecuador Venezuel a Brazil Regional data/app consolidation Workstation-less services Full performance and mobility Regulatory compliance for data U K Norwa y Algeria Nigeria Cong o Angol a It al y Hollan d Croati a Tunisia Li by a Egy pt Ir a n Kazakhstan Pakista n India Beijin g Indonesi a Australia
Thank you! 18 Andrea.Rodolico@nice-software.com Twitter: @nice_hpc