ComputaonalNeedsandHigh+ PerformanceCompungin+...
Transcript of ComputaonalNeedsandHigh+ PerformanceCompungin+...
![Page 1: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/1.jpg)
J U L Y , 2 0 1 3 | I E E E P E S G M
Eugene Litvinov
Vancouver, 2013
Computa(onal Needs and High Performance Compu(ng in Power System Opera(on and Planning
![Page 2: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/2.jpg)
Evolu(on of Compu(ng Power
© Franz Franchetti
2
![Page 3: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/3.jpg)
Evolu(on of Compu(ng Power
© Franz Franchetti
3
![Page 4: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/4.jpg)
Defini(on
• High-‐performance compuCng (HPC) is the use of super computers and parallel processing techniques for solving complex computaConal problems. HPC technology focuses on developing parallel processing algorithms and systems by incorporaCng both administraCon and parallel computaConal techniques.
4
![Page 5: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/5.jpg)
Why HPC?
• The size • The complexity of the power grid today and nearest future:
– Distributed resources – Renewable resources – High volaClity – Microgrids – Controls
• The Data • The evoluCon of compuCng – the cost of HPC is dropping
5
![Page 6: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/6.jpg)
Why HPC?
6
![Page 7: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/7.jpg)
Power Flow Model Today
7
![Page 8: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/8.jpg)
New Opportuni(es
• Current studies are always a compromise between confidence and Cme
• Large margins to account for: – Uncertainty – Not enough Cme to run all scenarios needed – No systemaCc way of selecCng scenarios
• HPC will allow for higher confidence and save costs on lowering margins
• ReformulaCon of convenConal tasks with new mathemaCcal models is impossible with tradiConal compuCng environments
8
![Page 9: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/9.jpg)
Three Primary Areas of Responsibility
Markets
Planning
Reliability
9
![Page 10: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/10.jpg)
Market and System Opera(on
10
Real-‐Cme OperaCon • System Monitoring and Alarming • AutomaCc GeneraCon Control • Load and Wind ForecasCng • Security Analysis • Economic Dispatch • Emergency OperaCon • Look-‐ahead Unit Commitment • TransacCon Scheduling • Intra-‐day Reliability Assessment
Long-‐term Planning
• Resource Adequacy • Transmission Planning • Strategic Planning
Mid-‐term Planning
• GeneraCon Maintenance Scheduling • Transmission Outage Scheduling
Short-‐term Planning
• Outage CoordinaCon • Load ForecasCng • Reliability Assessment • OperaCons Planning
Forward Capacity Market
• Forward Reserve Market • FTR Market
Day-‐ahead Energy Market
Real-‐Cme Market
Physical System Operation and Planning
Market Operation
![Page 11: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/11.jpg)
Possible Applica(ons
• Power System SimulaCon
• Long-‐term and strategic planning
• Short-‐term planning
• Market OperaCon
• Market SimulaCon
• Real Cme monitoring
• Control • R&D
11
![Page 12: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/12.jpg)
Power System Simula(on
• SimulaCon is a fundamental element in power grid operaCons
• Complexity: – Nonlinear, non-‐convex funcCons – Discrete and integer variables – Ill-‐behaved characterisCcs – Hundreds of Thousands of differenCal and algebraic equaCons
• Large volume of data • SimulaCon is slow and takes long Cme • Need high performance compuCng techniques and advanced compuCng hardware
12
![Page 13: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/13.jpg)
Transmission Planning
Reliability Planning:
• Long range Cme horizon (1 to 10 years)
• Transmission planning studies need to consider – Different transmission plan alternaCves/topologies – Different load levels (peak, shoulder and light load) – Different stress direcCons (generaCon dispatch) at each load level – Thousands of conCngencies (N-‐1) – 2nd level conCngencies N-‐1-‐1
• Steady state analysis: thermal and voltage analysis
13
![Page 14: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/14.jpg)
Transmission Planning
• Dynamic stability: Cme domain simulaCon – Need to model all dynamic devices (generators, exciters, governors,
power system stabilizer, HVDC, power electronics, relays, etc) – Tens of thousands of nonlinear ODEs with Cme steps in milliseconds
and over 50,000 nonlinear algebraic equaCons – One twenty seconds Cme domain simulaCon normally takes 20
minutes of CPU Cme
14
![Page 15: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/15.jpg)
Transmission Planning
• Southeast Massachusegs and Rhode Island assessment – Needs Assessment
• 36 power flow case, 295 first level conCngencies, 2122 second level conCngency
• One N-‐1-‐1 AC analysis takes about six minute • 36*295*6 = 63,720 minutes = 1,062 hours
– SoluCon Study: at least five different alternaCves
• Maine Power Reliability Program (MPRP) stability study – 11 power flow cases, 477 dynamic conCngencies – One twenty-‐second dynamic simulaCon takes about 20 minutes in
PSSE – 11*477*20 = 104,940 minutes = 1,749 hours
15
![Page 16: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/16.jpg)
Transmission Planning – Economic Planning
• Open stakeholder process to study alternaCve future system scenarios such as: – Large-‐scale integraCon of renewable resources in the 20-‐year Cmeframe
– Interregional analysis of increased transfer limits conducted by the Joint ISO/RTO Planning Commigee
– ReCrement of coal and oil units – Strategic studies – ProducCon cosCng
• Metrics: – Economics: ProducCon Cost, LSE Energy Cost, LMP, CongesCon – Environmental emissions (SO2, NOx, etc) – Fuel consumpCon / energy by fuel type
16
![Page 17: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/17.jpg)
Opera(ons Planning
• Shorter Cme horizon (up to 1 year)
• OperaCons planning studies need to consider – Line outage condiCon – Different load levels (peak, shoulder and light load) – Different stress condiCons (generaCon dispatch) at each load level – Hundreds of conCngencies (N-‐1)
• Close to two hundred stability operaCng guides
17
![Page 18: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/18.jpg)
Resource Adequacy
• ProbabilisCc study using Monte Carlo simulaCon to determine installed capacity requirements, Ce benefit, etc
• Typically users run 1000 replicaCon years, which takes around 10 hours
• Number of sensiCviCes
• Each replicaCon is independent and can be parallelized • Transmission Model
18
![Page 19: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/19.jpg)
The Benefits of HPC for Planning
• Produce more efficient soluCons
• Produce more robust soluCons
• Allow engineers to spend more Cme on “engineering” rather than data preparaCon, processing and waiCng.
• Allow solving problems that otherwise impossible
• More efficient use of the infrastructure
• Avoid compromise due to lack of Cme
19
![Page 20: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/20.jpg)
On-‐line Security Analysis
• Real Cme operaCon needs to calculate the system security and stability limits Cmely and accurately to enhance the operaConal decision making and reduce the risk of cascading failure
• On-‐line conCngency analysis, voltage stability analysis and dynamic stability analysis
• AutomaCc prevenCve control and correcCve control recommendaCons
• ConCngency screening and ranking techniques • Time domain simulaCon
20
![Page 21: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/21.jpg)
Power System Opera(on
• PrevenCve vs CorrecCve paradigm
• The need for superfast computaCon
• Faster than real Cme simulaCon
• Early warning systems
• Ability to use new controls • Decision support systems
• Online monitoring of system resiliency
• Wide area special protecCon systems
21
![Page 22: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/22.jpg)
Market Opera(on
• The need for high performance in Day Ahead market due to gas dependencies
• Use of stochasCc programming to deal with uncertainCes
• Market simulaCon
• MulC-‐stage decision support and pricing
• Market coordinaCon
22
![Page 23: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/23.jpg)
“Big Data”
• Explosion of data available at the control centers • Phasor Measurement Units (PMU)
• Data mining applicaCons for event detecCon
• Intelligent alarming
• NaCon-‐wide and regional repositories for situaConal awareness
23
![Page 24: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/24.jpg)
Supercomputers
• There is not enough staCsCcs to validate system planning results
• Need for massive amount of calculaCon to periodically validate assumpCons and engineering decisions
• May take years to run
• SimulaCon of future alternaCve grid architectures
• SimulaCon of interacCons with other infrastructures on naConal or regional scale
24
![Page 25: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/25.jpg)
• Why consider robust unit commitment (UC)? • Increasing number of renewable and demand response resources
• What is robust unit commitment? • Takes into account a range of the uncertainty, rather than a point
forecast as in the determinisCc reliability UC process. • Guarantees the feasibility of the system. • Requires mild informaCon of uncertainty, so no more guess of the
probability distribuCon.
• Is robust unit commitment beger? • Compare the robust approach with the determinisCc approach by
running a large number of Monte Carlo simulaCons. • Evaluate the operaConal and economic benefits • IdenCfy the opCmal conservaCsm level of the robust approach.
EvaluaCon of Robust Unit Commitment
25
load
hour
![Page 26: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/26.jpg)
Prior to the hpc4energy collaboraCon • Simulate the real-‐Cme dispatch process
• Codes in use – Implemented in GAMS using CPLEX 12.1.0 solver on a PC laptop with an Intel Core (TM) 2Duo 2.50GHz CPU and 3GB memory.
– Take about 30 minutes to solve a robust UC problem. – Take about 15 seconds to solve a dispatch problem in the simulaCon.
– 1600 robust UC runs × 1 M dispatch runs = 761 years
26
Solve Robust UC for 0, 0.1 , ... ,N NΔ = g
Solve DeterminisCc UC for
Robust UC0
Robust UCNd
…
Det UC0
Det UCNd
…
Simula(on: 1000 real-‐Cme dispatch
runs
SimulaCon
Measures: economic and operaConal performance
SimulaCon
SimulaCon
Measures
Measures
Measures … …
0, 0.1 , ... ,N NΔ = g
![Page 27: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/27.jpg)
Current Results
• Solve robust UC problems – MulCple robust problems can be solved at the same Cme. 1 robust UC run per core 10 cores per computer node 5 nodes per scenario 32 scenarios = 1600 robust UC runs
• Select scenarios for simulaCons – Number of historical data is small – Use the “cluster method” to create samples for simulaCons
27
• Time to solve 1600 robust UC problems decreased from 800 hrs to 30 minutes • Allow a more comprehensive evaluation
![Page 28: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/28.jpg)
Value added by LLNL hpc4energy collaboraCon
• HPC can efficiently run millions of simulaCons, which is impossible to be achieved by the compuCng capability of ISO NE.
• Allows more flexible evaluaCon designs. • SimulaCons yield more staCsCcally significant results.
• Enable ISO NE to perform more accurate and more comprehensive evaluaCon of the robust UC.
28
![Page 29: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/29.jpg)
Cloud
29
![Page 30: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/30.jpg)
Internal Grid Compu(ng Environment
• SimulaCon tool: – Siemens PSSE – PowerGem’s TARA – GE MARS
• Grid compuCng tool: – Enfuzion
• ConfiguraCon: – Control computer: one (psseroot) – Worker computers: four, a total of 40 processors – User computers: individual’s desktops
• Users need to prepare Enfuzion RUN file and submit for execuCon
30
![Page 31: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/31.jpg)
Issues with today’s Internal Environment
• Long wait Cme due to the increasing number of users and workload
• Scalability: – Maximum capacity of the system is reached – Adding addiConal resources takes Cme – Hard to esCmate the peak demand and average business compuCng needs
and purchase the IT infrastructure accordingly
31
![Page 32: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/32.jpg)
Objec(ves of the PoC Cloud Compu(ng Project
• Get hands-‐on experience and knowledge of the cloud compuCng technology
• Experiment deploying power system applicaCon in the cloud environment
• Benchmark performance between cloud run and internal run
• EsCmate cloud infrastructure usage cost
32
![Page 33: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/33.jpg)
Amazon EC2 Cloud Compu(ng
• Proof of Concept project with Cycle CompuCng • Amazon EC2 to provide the cluster grid • On-‐demand creaCon and use of high performance
compuCng environments as a metered, pay for use, service – Reserved instance pricing – On-‐demand instance pricing – Spot instance pricing
• Condor scheduler technology for load distribuCon and CycleServer for monitoring progress of workload and job scheduling
• EncrypCon schemes to secure data at various stages of the computaCon pipeline.
33
![Page 34: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/34.jpg)
Engineering Applica(on • TARA -‐ Transmission Adequacy & Reliability Assessment
• Sotware registraCon issue – Special cloud license
• Engineering study – AC N-‐1-‐1 conCngency analysis – Twenty five power flow cases with 164 line-‐outs, a total of 4100
combinaCons (jobs)
34
![Page 35: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/35.jpg)
Test 1
• Smaller power flow case (5MB) with equivalent of all areas outside of NPCC – > 340 hours on individual’s desktop – > 8 hours over the internal Enfuzion environment
• Amazon EC2 – C1.xlarge instance: 8 cores, 7GB RAM
– Setup Cme of compuCng nodes: 10~20 minutes
Cost: – 20 * 3 hours * $0.20/hr = $ 12 – 50 * 2 hours * $0.20/hr = $ 20 – 100 * 2 hours * $0.20/hr = $ 40
35
![Page 36: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/36.jpg)
Test 2 & 3 • Full power flow case without equivalent (23 MB)
– > 1700 hours on individual’s laptop – > 40 hours over the internal Enfuzion environment
• Amazon EC2 – C1.xlarge instance: 8 cores, 7GB RAM – 150 compuCng nodes – 90 minutes to finish with cost about $60
• EncrypCon enabled (AES-‐256) – S3 server side encrypCon on all data transfer to and from S3 storage – EncrypCon on shared file system between Linux scheduler and windows execuCon
nodes
• The same case as in test 2 was run in test 3
• No performance impact was noCced due to encrypCon
36
![Page 37: ComputaonalNeedsandHigh+ PerformanceCompungin+ ...ewh.ieee.org/cmte/pes/etcc/E_Litvinov_HPC_Power... · JULY,2013|IEEEPESGM Eugene%Litvinov% Vancouver,+2013+ ComputaonalNeedsandHigh+](https://reader036.fdocuments.us/reader036/viewer/2022071117/6003676adfed6d63054f56fd/html5/thumbnails/37.jpg)
Challenges in Adop(ng HPC
• Besides “embarrassingly” parallel applicaCons (like conCngency analysis), current sotware is not created for HPC.
• Major computaConal engines have to be rewrigen for parallel execuCon
• New class of decomposiCon techniques for solving opCmizaCon problems, ADE, etc.
• New class of algorithms and sotware has to be developed for using new technologies like GPU
• Vendor licensing
37