Internal:CNF HPC Workshop 2019

Presentation Files

https://odu.box.com/s/6az46pa7zr9jhf1s05w6gqrs8lm4u5o5

Cara Leckey

• High performance incubator program (HPCI)
• Nasa has projects involving both earth sciences and sun-/other-planet-based sciences
• check provided video, contains reserach at all levels, like design- implementaion and validation
• Exploring science from Moon to Mars
• Software at NASA has long lifespan, often is 20+ years
• HPCI focuses on building the skills for effient and maintinable software
• Computatubal nodestrictiove evaluation, : simulations for defetc detection aircraaft and material focuded ultrasound simulations
• EDL : entry descent and landing project for mars and other planets
• MLMCPy replacing putre model carlo gives them about 10x improvement

Amber

• Buikd arround CEBAF ( Continiuoud electron beam accelator)
• unique in terms of functionality ( uses electrons for probing)
• Operational since 1995, 1630 active users !
• 1/3 of students in nuclear physics come through JLAB
• spinoff techonologies on medical field ** imaging ** cryogenics
• shipy cluster (2016 was on top 500) ** 17k cores 12,5k cores mainly for QCD (quantum chromodynamics) not analytically solvable need numberical solution !
• Jlab needs not only offsite but also onsite powerfull computing facilities to handle strewming data
• use simualted data sources to complement experimental like NASA does!

David

• chromo is chromodynamics has to do with the “colours” for the quarks
• you nwcwr see quarks alone, the force between in them is extrimly high
• QED/QCD
• Big question is how are quarks and gluons distributed inside the nucleon
• Jlab new developement ny times how big is the proton
• Electron ion has the difference that we do not collide an electonr with a stiic target but with aa proton beam allowing thus higher speeds
• Like in medical imaging we try to rebuild image fromsparce and incomplete data
• expand to other areas , like visualizarion for understading , verification and debugging
• essential we are solving an inverse problem like in the black hole imaging case

Eric

• about 10 main CFD codes, that was the speakers background
• CFD vision 2030 document ** advocational document ** practices for nationalwide CFD analyses ** used everywhere where funding matters ** PI was Boeing
• turbulent flows and separation is the main focus
• cruse speed, drag , lift easy already
• separaion and turbulent is the different, landing and takeoff for example
• or how Boeing can be still competitive
• High lift is the most difficult , close to 2 billion for certification of a new boeing aricraft
• ie during landing , landng gear and flaps open
• Big impact
• more than 10k simulation before anything is cut at Boeing
• Capacity job was the norm in order to build databases of small scalae solution
• But capability is need in order to advance
• RANS no enough swithcinh to hybrid RANS/LES
• How to do mesh adaptation on 1000 cores?
• How to hangle geometry ?
• Retropropulsion simulations, ** about two-store vehicle , how to land it on mars ? ** biggest until know was mini-couper size ** landing withing 5 min from landing speed to touchdown
• typical 6 billion element, 200k time spes , 200TB per run
• see number for convetional ** 3 months run on 5000cores 6 months in queue
• now on summit : ** 4 days on 552 GPUS equi to 100000SKL cores ** 30-40 TB per day

Ed

• material vision 2040 in the same realm like CFD 2020
• NASA/CR-2018-219771
• wide spectrum of scaling nano-micro to cm
• assertenty verification
• understant micro-structures
• save mass and material without loosing interity structure
• pipeline includes from microstuturess to localized FEA models and contruction integrity

Pete

• ATPESC (best tra
• complicating heteronegeious memory layout
• everything is dynamic
• AML (memory libary)
• Argo library *AML on chip memory, easy way to use it as a cache but we can do better than that dive in and write code in RTS to handle it readthedocs page ! argo-aml.readthedocks.ion
• change shape of data structure during transfer through DMA
• futute architectures will be heteregoneous we need to eb able to move data arround
• Ken batcher , supercomputer is a a device for turning compute-bound problesm in to I/O bound problem
• Edge computing is the opposte
• move some of the computtions to edge to reduce data output
• used by hydrology, national secutiry
• new companies comming up with new specialized hardware
• Google edge GPU great but only for one specific thing Tensorflow
• Sage prject , spread accross USA, wild fire prevention
• do machine leanring at the edge to singl if smoke is seen
• never send raw data
• NSF NEON, capture and integrate data @ Miltiple scales
• XLOOP workshop
• How about Japan with many arm based nodes ** extexted arm for vectors ** very power efficent

Nikolopoulos Dimitris

• use optimism to improve your algorithms
• how to do now optimistics computing going beyond what computer hardware allows you to do
• How tp=o improve and optimize all the time for all kinds of edge computing
• Can we deliver high-performance cpomputing depedable computing under unsertainly
• TAPAS joined work with IBM. Train-less Neural Network Architecture Exploration tool
• maybe you do not have the time or resouaches to do time consuming training
• What if instead of training we teach how to learn ????
• Learn how networks learn instead of training a network. Find a way to encode them quickly.
• TAPAS goes very quickly over more than 1000 networks/se with low end GPU
• Already deployed in contraines and microservece
• Threes steps ** Data charecterization aginast difficulty (quick) ( how a base probe network corresponds to the dataset ?) ** Build knowledge base and check if anything seen in the past has performed well in these diffuculty ** Search the db.
• Ecnode type and connectivity of nodes as well as the resource requirements
• DEFCON-CPU ** CPUS can run up to 25% faster but finding the right amount of fiddling is diffucult once you go above you have timing issues ** applied on real hardware on ARM processors

1. identity workload-specific behavior, timing-error-prone code sequences
2. Mitigate those by building a voltage and frequency governorns

• Firsth they found sequncies that cause timing errors and then they identifies these sequncies in real-world problems
• HERMES ** same idea but for memory ** relax hardware limits and find the errors *DRAM fereshes evey few ns to keep data alive, but you also refresh on write so can we minimize refresehes ?
• build on top of linux NUMA module to Reliability data regions
• can save a lof of memory
• summary ** use compiler like optimization back to machine leanrning