EUROEXA CO-DESIGN, DEMONSTRATION AND EVALUATION USING A RICH SET OF EXASCALE-CLASS APPLICATIONS - ESIWACE
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
EuroEXA Co-design, demonstration and evaluation using a rich set of exascale-class applications Graham Riley & Mike Ashworth, Andrew A1wood, John Mawer Advanced Processor Technologies: apt.cs.manchester.ac.uk School of Computer Science, University of Manchester graham.riley@manchester.ac.uk © 2017–2018 EuroEXA and Consortia Member Rights Holders Project ID: 754337
EuroEXA • Build on foundaFon from previous projects • Co-design and demonstrate 1 PF+ testbed in operaFonal environment • Apply design flexibility of UNIMEM architecture • Aim to deliver available performance to full-scale producFon applicaFons • Achieve balance between compute resources and demands of applicaFons • Technology is ARM mulFcore + FPGA for acceleraFon of apps and comms 23 Jan 2018 2
What is EuroEXA? AddiFonal IP Open CompeFFve Exercise Other Partners Organised by the EU EU Funding & Monitoring Alliance of MulFple Projects, IP and Partners 23 Jan 2018 3
EuroEXA partnership Commercial Partners Academic/Gov. Partners Supporters • September 2017 – February 2021 (3.5 years) • Budget €20 M 23 Jan 2018 4
System architecture and technology: Compute node Technology from FORTH (QFDB board): • 12 cm x 13 cm Censored Censored • 4 ARM Processors and 4 FPGA Accelerators • M.2 SSD Censored Censored • 4 x SODIMMs + Onboard RAM • Daughterboard style • 160 Gb/s of I/O 23 Jan 2018 5
System architecture and technology: Blades Technology from Iceotope: • 16 Node half depth 1u chassis • 2 x 3.2 kW per u (back2back) • Total Liquid Cooling technology • 48 V DC distribuFon • Hot water out, chiller-less operaFon 23 Jan 2018 7
2.5m 2MWICEOTOPE CONFIDENTIAL Aisle of Load of Load Batteries and Rectifiers for 250kW Batteries and Rectifiers for 250kW Wall 600x600mm 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs 600x600mm 600x600mm cabinet with 6 cabinet with 6 Aisle chassis & 2x CDUs chassis & 2x CDUs Coolant and DC Power Coolant and DC Power 600x600mm Technology from Iceotope: 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs 600x600mm 600x600mm • 2MW per shipping container cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs 600x600mm 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs Aisle 10.6m 600x600mm 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs 600x600mm 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs Project ID: 754337 600x600mm 600x600mm cabinet with 6 cabinet with 6 • Enables 3D stacking to minimise distances Aisle chassis & 2x CDUs chassis & 2x CDUs 600x600mm 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs Coolant and DC Power Coolant and DC Power • Shipping Container OpFmised Cabinets (2 rows) 600x600mm 600x600mm cabinet with 6 cabinet with 6 chassis & 2x CDUs chassis & 2x CDUs Wall © 2017 EuroEXA and Consortia Member Rights Holders Aisle Aisle of Load of Load Batteries and Rectifiers for 250kW Batteries and Rectifiers for 250kW 10mx2.4m In 40) ISO or Technology – Data Centre w/index.php? curid=29173440 By Hubert.Knop - Own commons.wikimedia.org/ work, CC BY-SA 3.0, h1ps://
Our path: prototypes • Three testbeds to be deployed at STFC, Daresbury Testbed 1 Testbed 2 Testbed 3 50 nodes of new 500 nodes and new Test new node and technology for infrastructure processor technologies so)ware technologies to test that will ulFmately development scaling project to Exascale Mid 2018 Late 2018/ 2020 early 2019 23 Jan 2018 10
EuroEXA: co-design, demonstration and evaluation using exascale-class apps X V chem 2D chem 3D x! U physche m FLOPS L1000 HCI Y AVU-GSR Quantum Espresso SMURFF NeuromarkeFng IOPS Mem BW InfOli Astronomy FRTM NEMO image classificaFon NEST/DPSNN Mem capacity IFS Alya LFRic LBM GADGET 23 Jan 2018 11
Why these 14 applications? • Covers three important applicaFon domains • Climate and weather (LFRic, NEMO, IFS) • Physics and energy (LBM, Alya, GADGET, AVU-GSR, FRTM, Astronomy image classificaFon) • Life science and bioinformaFcs (NEST, NeuromarkeFng, InfOli, Quantum E, SMURFF) • All domains will require exascale compuFng in near future • ALYA, GADGET, NEMO and Quantum ESPRESSO part of PRACE UEABS 23 Jan 2018 12
EuroEXA programming environment • ApplicaFons will target portable programming models / communicaFon libraries • MPI, GPI (GAS programming Interface), Vivado HLS, OmpSs@FPGA, MaxJava, OpenCL (via SDSoC/SDAccel), OpenStream, and PSyClone • OpFmizaFon for EuroEXA architecture done by programming environment • Building on the work in ExaNoDe project • Greater maturity, especially in FPGA support • Tuning and opFmizaFon through EuroEXA’s full applicaFons (cf ExaNoDe mini-applicaFons) • Though work in Weather and Climate is focussed on kernels/mini-apps/dwarves currently • Partners are currently porFng to x86 clusters and Trenz boards / UltraScale+ • Will migrate to Testbed 1, 2 and 3 as available 23 Jan 2018 13
EuroExa software stack… Building on ExaNoDe ExaNoDe and ApplicaFons EuroEXA EuroEXA Accelerated libraries Hardware Psyclone Maxeler tools Nanos OpenStream SLURM Resource BeeGFS MPI (MPICH and OFI) GPI allocaFon Hyperconver opFmizaFon ged storage NaFve UNIMEM API, device drivers, and firmware FPGA for compute FPGA for UNIMEM ARM cores 23 Jan 2018 14
Conclusions and vision • First testbed architecture will be shown to be capable of scaling to world-class peak performance in excess of 400 PFLOPS with an esFmated system power of around 30 MW peak • Co-design, demonstrate and evaluate using rich set of exascale-class applicaMons • Demonstrate / make progress on FPGA-enabled applicaFons • Specifically, for Weather and Climate, looking at LFRic, NEMO and IFS 23 Jan 2018 15
You can also read