Transcript technology_module1 - NDSU Computer Science

HPC Technology Track:
Foundations of Computational Science
Lecture 1
Dr. Greg Wettstein, Ph.D.
Research Support Group Leader
Division of Information Technology
Adjunct Professor
Department of Computer Science
North Dakota State University
What is Computational Science?


Application of computational methods to
fundamental problems in science and engineering.
Concerned with application of computational
methods to one of three objectives.

Modeling

Simulation

Prediction
Modeling


Involves development of a mathematical model
capable of predicting physical phenomenon.

weather prediction

molecular energy and force fields

properties of materials
Typically involves solving systems of differential
equations expressed in linear form.
Challenges in Modeling


Currently only order 0 systems can be solved
directly.
Computational approaches are frequently
implemented with methods involving truncated
series.



Garbage In/Garbage out (GIGO)
Limitations constantly demand attention to model
parameterization and viability.
The field considers better differential equation
solvers to be essential.
Emerging Opportunities in Modeling


Model parameter evaluation.

Considered important by the national laboratories

Involves computation of first and second
derivatives of the model with respect to finalized
model parameters
Important focus is to determine quality or stability of
model.
** Model evaluation is critical.
Multi-Scale Studies


Defn: Calculation of system behavior or properties
on one level using information from subordinate
levels.
Continuum of levels (physical systems):

quantum mechanical

molecular dynamics

meso or nanoscale levels

level of continuum

level of device
Simulation



Focuses on simulating the behavior of physical
systems.
Usually involves Monte Carlo methods to solve
stochastic systems.
Most commonly employed in computational
physics.
Simulation – con't.



Central to the 'birth' of computational science.

Metropolis, Rosenbluth, Rosenbluth, Teller and
Teller

“Equation of State Calculations by Fast Computing
Machines”
Goal is to develop 'ensembles' or collections of
parameters.
Typically implemented as 'coarse grained'
parallelism.
Prediction




The analysis or 'mining' of large sets of data for the
purpose of predicting future phenomenon.
Centrally important to marketing and e-commerce.
Represents a type of computational problem referred
to as 'embarassingly' parallel.
Most famous example is NetFlix competition.
Challenges in Prediction



Data locality

More processors does not equal more speed.

NetFlix competition demonstrated inadequacy of
improperly 'balanced' computational architectures.
Primary concern of national labs involved in security
based computation.
Current HPC architectures exacerbate data locality
problems.
'The Wettstein Rule of
Computational Reality'
“If filling a cache line is too slow you will
be really unhappy doing a cross-node
lookup to a machine 200 racks away.”
RoadRunner Configuration

1.72 petaflops peak / 1.456 petaflops demonstrated

296 racks covering 6,000 square feet

Massively parallel – hybrid architecture


6,480 Opteron (x86) processors

12,960 IBM PowerXCell processors

122,400 cores
103.6 terrabytes of memory
But when is it fast?
A Tradeoff

Compelling speeds when each node can work on a
discrete element of the problem.


Embarrassingly parallel problems.


Strictly orthogonal decomposition
MIMD
Less efficient when.

Boundary condition dependent problem.

Access to entire memory space is required.
Latency
The Enemy of Prediction

Latency definition:



The amount of time required to retrieve the next
relevant item of data required in a computational or
predictive sequence.
'Achilles Heel' of modern massively parallel systems
such as RoadRunner.
Common problem since the design of the Cray-1.

Wiring optimized to place time critical connections
on the inner portion of the computer.
Reducing Latencies through PTree's



Current area of research interest.
Addition of second order PTree's to optimize data
selection decisions.
Minimizes:

cache line flushing

cross-node data lookups
Exercise

Log into cluster1.chpc.ndsu.nodak.edu.

Use sinfo command to locate an available node.


Use ping command to measure message latency over
standard TCP/IP network.


e.g. node64-49
ping -c 5 node64-49
Use ping command to measure message latency over
Myrinet:

ping -c 5 node64M-49
Exercise – con't.


Bottom of ping command details min, average and
maximum communication latencies.
Compute expected performance change if a
computation is constrained by the length of time
required to pass a message from one node to another.
goto technology_lecture1_2;