optimizing for the exascale era - red hat...nov 21, 2019 · –epcc: wrf, openfoam, rolls royce...
TRANSCRIPT
Optimizing for the Exascale Era
COMPANY FOUNDED
1995FY19 REVENUE
$2.9BEMPLOYEES
5,000+
LOCATED IN
Santa Clara, CA
R&D CENTERS
US, Israel, India,
Germany, China
PATENTS WORLDWIDE
10,000+
2
Marvell Overview
© 2019 Marvell. All Rights Reserved.
STORAGEPROCESSOR, SECURITY,ARTIFICIAL INTELLIGENCE
© 2019 Marvell. All Rights Reserved. 3
Delivering the essential building blocks for infrastructure
NETWORKING
Artificial IntelligenceAccelerators & offload processors
AutomotiveSecure Ethernet, PHYs, storage & Wi-Fi
Data Center Storage SolutionsNVMe aggregators, accelerators & converters
Storage ControllersHDD & SSD
#1 in HDD and SSD controllers
Fibre ChannelAdapters & controllers
#1 in Fibre Channel │ 20M+ ports shipped
Wireless ConnectivityWi-Fi access points & clients
Industry’s Best Wi-Fi 6 solution
NetworkingEthernet switches, PHYs & NICs
#2 in Networking │ 3M+ FastLinQ ports shipped
ComputingServer, baseband & embedded processors
#1 in baseband and multi-core processors
SecurityFIPS & virtual offload
#1 in security processors │ 1M+ LiquidIO ports shipped
© 2019 Marvell. All Rights Reserved. 4
Marvell has a proven track record in processors
Billions of CPUs shipped
15+ years of high-performance multi-core CPU expertise
Multiple generations of Arm®-based server CPUs
Arm architecture tremendous foundation for server design
• Modern 64-bit architecture
– No legacy
– Smaller core, better power
• Custom designed for data centers
– No common core design for desktops/laptops/servers
– Area & power advantages
– Single core/thread performance now comparable to x86
• Maturing ecosystem
• New Arm native use cases at the cloud and edge
• Open architecture allows competition & DIY
5
© 2019 Marvell Confidential, All Rights Reserved.
© 2019 Marvell. All Rights Reserved.
6
Arm servers well-suited for data center applications
• Better throughput
• Improved application latency
• Arm applications natively run in the cloud
• Better power
Cloud compute Arm-native/AndroidTM Edge computingBig data/HPC
© 2019 Marvell. All Rights Reserved.
ThunderX2: 2nd generation Armv8-A based server processor
• Up to 32 custom Armv8.1 cores, up to 2.5GHz
• Full OoO, 1, 2, 4 threads per core
• 1S and 2S configuration
• Up to 8 DDR4-2667 memory controllers, 1 & 2 DPC
• Up to 56 lanes of PCIe Gen3, 14 PCIe controllers
• Full SoC: Integrated SATAv3 USB3 and GPIOs
• Server class RAS & virtualization
• Extensive power management
• LGA and BGA for most flexibility
• 40+ SKUs
7
Production: 2018
© 2019 Marvell. All Rights Reserved.
8
ThunderX2: Most widely-deployed Arm-based server processor
© 2019 Marvell. All Rights Reserved.
Mission & System Overview
Catalyst: Accelerating Arm Adoption in UK
12,000 ThunderX2 cores across 3 sitesAmong the largest Arm HPC global deployments
Industry Partners
Mission
– Early adoption of Arm for HPC in UK
Goals
– Port/demonstrate critical HPC apps,
engage with ISV, & enable follow-on
collaborations
– Fix and file bugs & up-stream to open
source
System Overview
– Processor: 32-core 2.2 GHz Marvell
ThunderX2 Arm processor
– Memory: 256GB (16 x 16GB DR
DIMMS)
– Service & storage: AR64z, 180TB
(SSD/HDD)
– HPE: Insight CMU
– Arm: Allinea Compiler, Libraries, DDT, MAP
– Mellanox: OFED, HPC-X, OpenMPI, OpenSHMEM,
SHArP
– SuSE: SLES, OpenStack, OpenHPC
– Marvell: Thunder-X2 SoC, technical support
UK Collaborations
– UK Government: Dept. for Bus., Energy &
Industrial Strategy (BEIS)
– EPCC: WRF, OpenFOAM, Rolls Royce
Hydra opt, 2 PhD candidates
– Leicester: Data-intensive apps, genomics,
MOAB Torque, DiRAC collab
– Bristol: VASP, CASTEP, Gromacs, CP2K,
Unified Model, Hydra, NAMD, Oasis, NEMO,
OpenIFS, CASINO, LAMMPS
73 HPE Apollo 70 Dual Socket
ThunderX2 arm processor
servers per site assembled in
2 racks
Each deployment contains 66
AR44z (1U) & 7 AR64z (2U)
HPE Apollo 70 server nodes
https://news.hpe.com/academia-and-industry-collaborate-to-drive-uk-supercomputer-adoption/
https://arxiv.org/pdf/1904.04250.pdf
Catalyst blog link9
IHVs
10
ThunderX2: Most widely-supported Arm server processor
Applications
and Tools
OEM and ODM
platforms
OS / Firmware
Middleware
EDGECLOUDHP
Over 100 partners across commercial, open source, and industry standards engagements
© 2019 Marvell. All Rights Reserved.
© 2019 Marvell. All Rights Reserved.
ThunderX2 and GPU Optimization with NVIDIA
HPE Apollo 70 AR64z ServerTwo GPUs per server, two servers per Chassis
• Initially announced at ISC 2019
• CUDA-X AI™ and HPC libraries
• PGI compiler included
• Early applications available now
11
• Close engineering collaboration on kernel and OS development.
• 2015: RHEL 7 for Arm Development Preview• 2017: RHEL 7 for Arm• 2019: RHEL 8 includes Arm support
• Commercial support on HPE Apollo 70 and Bull Atos.
• RHEL 7.x and 8.x is available from Red Hat, including free Red Hat Developer subscription.
• Joint industry engagements in standards, Arm, Linaro, and others.
• Community engagements in Fedora and CentOS.
© 2019 Marvell Confidential, All Rights Reserved. 12
Red Hat & Marvell Collaborationhttps://www.redhat.com/en/about/press-releases/red-hat-enterprise-linux-it-industrys-largest-commercial-open-source-software-ecosystem
© 2019 Marvell. All Rights Reserved. 13
Marvell ThunderX at SC’19
“NVIDIA CUDA-X AI and HPC Software Stack Now Available on Marvell ThunderX Platforms”
“Marvell Receives Honors in 2019 HPCwireReaders’ and Editors’ Choice Awards”
Marvell Ecosystem Partner BoothsAltair
Arm
DDN
Nvidia
Penguin
OSS
E4
Gigabyte
HPE
Red Hat
SUSE
Verne Global
Expanding investment into server processor roadmap
• Beyond direct Marvell investment, our customers and partners are investing directly into our business
14© 2019 Marvell. All Rights Reserved.
2016 2018 2020
ThunderX roadmapDriving >2X generational performance improvement
Production
Production
Coming soon
Stay tuned
7nm
16nm
28nm
15© 2019 Marvell. All Rights Reserved.