Performing Simulation-Based, Real-time Decision Making with Cloud HPC

Performing simulation based, real time decision making with cloud HPC Zack Smocha, April 2016

7 Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances

Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances2

Agenda

•  Rescale overview •  Evolution of simulation •  Simulation in service •  F1 simulations •  Manor Racing case study


HQ San Francisco, USA , Japan office rapid growth

Global simulation cloud HPC platform 30+ data centers, 120 simulation software

Over 100 leading enterprises - automotive, aerospace, energy and life sciences

Headquarters

Technology

Customers

Investors

Rescale - Company Overview

Peter Thiel Jeff Bezos Richard Branson

... and several other industry leaders, technology experts, and experienced executives

Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances47 Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances4

Rescale Cloud HPC Enterprise Simulation Platform

So#ware

120+ software packages

Mul,-clouds

30 varied location and HW availability

Workflow

Administra,on

Security

Compliant, data and user

Manage usage access and cost

Experienced team seamless workflow

5

Simulations in Industry

Aero

spac

e

Auto

mot

ive

Life

Scien

ces

Oil &

Gas

Indu

stria

ls

Sem

icond

ucto

r

•  Complexturbine•  Wingdesigns•  Modelling

propulsion

•  Crashsimula-on•  Engine

computa-onalfluiddynamics

•  Reservoirsimula-onworkflows

•  Hydrocarbontraps

•  Gene-cengineering

•  Isola-onofgene-ctraits

Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances


Evolution of Automotive Simulation Co

mpl

exity

1960 1970 1980 1990 2000 2010 Today

Vehicle Dynamics

Crash Analysis

FEA

Multiphysics High-fidelity Ensemble analysis

CFD & HPC


Simulation in the Product Life Cycle

Predict behavior without actually testing it in real life

Validate and optimize the design of parts and manufacturing

Using real data to help make real time decision

Engineering Design Manufacturing In Service/Production

Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances87 Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances8

Simulation In Service - Time Spectrum

Real-me“nextmove”gaminganalysis

Threads CPU GPU ELO

24 764 112 3,079

40 1,202 176 3,140

64 1,920 280 3,168

Maintenanceandabnormalbehavior

Usingreal-metracksidedataforracestrategy

Makesureyoudon’tcrashthebus

Days/hours Minutes MicrosecondsSeconds


F1 the Art of Race Strategy


F1 Results Australia - Many Strategies

Superso# So# MediumHard Wet Intermedium Used=PitStop

•  BasedinBanburyUK•  PartnerswithMercedes-Benzenginetechnology•  WilliamsAdvancedEngineeringfortransmissions


Manor Racing


Manor use case - Goals

•  Best-metotakeapitstop•  What-restofitforthenextstageoftherace.•  Secondguessthecompe--ontotryandgainraceposi-onthroughbe^erpitstop-me

For Manor Racing it is about meticulous attention to detail, eking out every single opportunity to find every single gap. Car and driver, factory and team


Manor use case - Users

DaveRyan-RacingDirector

JamesKnapton-HeadofVehicleScience

Strategyengineerswhoadvisetheraceengineersontheop-mumstrategyastheraceisdeveloping


HPC Simulation in Service - Requirements

•  Collect the data in real time? •  Insert the data into the system? •  Upload the data to the cloud HPC? •  Best HW for fast simulation? •  Download and access the data?


Manor Cloud HPC Architecture

Cloud HPC Cluster

Head NodeHPC Scheduler

Compute NodesNodes are joined to the HPC Scheduler

Virtual Network LAN

IPSec VPN

Manor application GUI

•  For optimization jobs directly interact with the HPC cluster •  Clients running jobs join the head node domain and mount the shared file system


Input Parameters and Live Data

•  Parameters: lap time, tire degradation rate for each tire compound, expected car performance as fuel level reduces

Make a live decision based on the

simulation results and enter actual track

side results

Collect live track side data and run the

simulation

Make a live decision based on the

simulation results and enter actual track

side results

Collect live track side data and run the

simulation

•  Example of live Input data: Actual lap time, tire degradation


Input Parameters

•  How do I collect the data input in real time – Data is available from the track side

•  Insert the data to the system – User enters the data into the Manor

application interface, application generates input size files ~500kB

•  Upload the data – Data is uploaded to the head cluster node

from the user laptop


Simulation Benchmark - Best HW for Fast Simulation

•  Simulations based on Monte-Carlo methods •  Response < 45-50 sec •  Run thousands of race simulations per minute, repeat

this process over and over throughout the race

#cars #cores Strategies Permuta-ons Itera-ons Running-meonthecluster

1Car 500 30 100 100 32.27

1Car 500 30 300 20 69.58

1Car 500 30 150 20 31.69

1Car 500 90 20 20 31.82

1Car 500 90 20 100 35.61

1Car 750 30 100 100 31.80

1Car 1500 30 100 100 30.75

2Cars 750(each) 30 100 100 35.03


Running the Simulation

•  Clusters are running the whole race

•  3000 tasks •  Hardware:

•  1500 cores •  16 CPU per node •  98 nodes •  Haswell CPU


The Results

•  Output: Optimum race time •  Results size is 5MB •  Results are download to the user PC •  User views results in the Manor GUI App •  Using the results in practice: Decide when would be the

best time for a pit stop


The Results

•  Each curve is a different tire choices •  Each # represent a pit stop and the lap to stop •  POA: Prime/Option/Alternate : Hardest to the softest

Op,mumrace,me–Sensi,vityto,restrategy

Rescaleconfiden-al–pleasedonotdistributeunderanycircumstances

Simulation defines our future, join us in helping build a better world.


Simulations in F1

•  Windtunnel,aerodynamic•  CFDandFEA•  2014FIAregula-ons


F1 the Art of Tire Change

•  15-19people•  Stopbelow3sec•  Dootheradjustments

Appendix
