hpc at saudi aramco’sarchive.hpcsaudi.org/events/2011_khobar/presentations/hpc at saudi... ·...
TRANSCRIPT
SAEED Al-ZAHRANI
HPC at Saudi Aramco’s EXPEC Computer Center
Agenda
Introduction to Saudi Aramco and ECC
Why do we use HPC?
ECC HPC Platforms
ECC HPC Strategy
ECC HPC Challenges
Saudi Aramco
Manages proven crude oil reserves of
260.1 Billion Barrels
4th largest gas reserves of
279 Trillion Cubic ft
Worldwide refining
capacity of
over 4 million BPD
Owns the world’s 2nd largest fleet of tankers
Fully integrated global petroleum enterprise
– Exploration,
– Producing,
– Refining,
– Petrochemicals,
– Distribution,
– International Shipping , and
– Marketing
http://www.saudiaramco.com
ECC Vision World-class
High Performance Computing
Data Center in Oil & Gas
ECC Mission • Support Upstream
• Scientific Computing
• Cost Effective
• Certified Professionals
• World-class Practices
EXPEC Computer Center (ECC)
EXPEC Computer Center
ECC Supporting Services
Why do we use HPC?
Main Functions – Seismic Processing
– Reservoir Simulation
Seismic Processing
• Conventional Processing
• Pre-Stack Time Migration (PSTM)
• Pre-Stack Depth Migration (PSDM)
Seismic Processing
Reservoir Simulation
An area of reservoir engineering in which computer models are used to describe fluids (oil/water/gas) and its flow dynamics in the subsurface geologic formations.
Gas
Oil
Water
Reservoir Simulation
• Understand fluid flow behavior
• Develop depletion strategies & forecast future
• Software
POWERS (Parallel Oil WatER Simulator)
GigaPOWERS • Can simulate billions of cells
• Giga models for super-giant fields
Low Resolution modeling
High Resolution modeling
Technology Evolution
1
100
10,000
1,000,000
19951996
19971998
19992000
20012002
20032004
20052006
20072008
2009
2010
2011
IBM Main Frame, CRAY IBM SP PC CLUSTERS
ECC HPC platforms
Driving Force: Do more with less!
Cray T90
IBM SP
PC Cluster (2002)
PC Cluster (2011)
12 Processors
512 Processors
1900 Processors
65,000 Processors
Impact of affordable HPC
• Dramatic improvements in price/performance bring out more computationally intensive applications that were previously thought to be cost prohibitive
• IBM SP: Enabled PSTM, POWERS
• Clusters: Full PSTM, PSDM, GigaPOWERS
• Heterogeneous (CPU + Accelerator/GPU)
– for RTM, full wave equation
Current HPC Environment
• 500 Teraflops total compute capacity
– > 6800 nodes, 65,000 cores, 97 TB RAM
• 43 HPC Clusters
– Different sizes & configurations
– 2 systems in the latest Top500 list
• Several Petabytes of high performance disk
HPC System Design
Clustered File System
(CSSP) I/O
Nodes
Disk Storage
Cluster A
Cluster B
Cluster Z
Co
mp
ute
N
od
es
MPI
CSSP (Conventional Serial Seismic Processing)
• Small number of nodes connected to High Performance Storage pool
• Clustered File System
PSTM (Pre-Stack Time Migration)
• Embarrassingly Parallel
• Master node on CSSP cluster
• Data fed via MPI to Grid of compute nodes
• GigE interconnect sufficient
HPC System Design
POWERS/GigaPOWERS
• Memory bandwidth is key
• Cannot use all cores in CPU due to this
• Needs low latency, high bandwidth interconnect
• Quadrics Myrinet Infiniband
• Typical clusters are 512 nodes
• 2 CPUs (12 cores) per node
HPC Strategy
• Meet current and future growth Scalable
• Expand and provide computing capacity as needed, avoiding unnecessary pre-investment. Modular
• Exploit the utilization of inexpensive computing resources that work together, to address complex problems that would otherwise require very expensive systems to solve.
Parallel
Key Factors
Meet Capability requirements
Reliability
Maturity of Technology
Ease of Management
Price / Performance
Sustained improvements in Roadmap
Trends
• Higher Resolution Seismic surveys
• Intelligent Fields (i-field) Huge increase in data
• Interactive visualization Giga-size models for higher resolution
• Minimize the cycle time Optimize decisions
• Increased complexity computing power Better Algorithms
HPC Challenges
•GPUs, Intel MIC (Many Integrated Core)
•Coding & Optimization is non-trivial
Heterogeneous Computing
•Memory bandwidth
•Interconnect bandwidth and latency I/O bottlenecks
•Data Management
•Data Backup/Archival Big Data
•Security
•Solutions Maturity
•Large Data transfer HPC in the Cloud
•Watching and Learning Design Challenges for Exascale
Summary
• Saudi Aramco leverages innovative technologies to deliver the best price/performance in HPC
• Constant efficiency Improvements in hardware, software and workflow
• Continual evaluation of alternate technologies