Transcript Introduction - Computer Science @ The College of Staten Island

Lecture 1 Introduction
Advanced High Performance
Computing
Fall 2014
1
Contents

Acknowledgments for today’s lecture
• Jack Dongarra (U. Tennessee) --- CS 594 slides from Spring 2008 —
http://www.cs.utk.edu/%7Edongarra/WEB-PAGES/cs5942008.htm
• Kathy Yelick (UC Berkeley) --- CS 267 slides from Spring 2007 —
http://www.eecs.berkeley.edu/~yelick/cs267_sp07/lectures
• Slides accompanying course textbook —http://wwwusers.cs.umn.edu/~karypis/parbook/
• Vivek Sarkar(Rice University) –
http://www.owlnet.rice.edu/~comp422/lecture-notes/comp422lec1-s08-v1.pdf
• Alexandros Gerbessiotis (New Jersey Institute of Technology)
2
Why parallel computing?
– computational modeling and simulation
“Computational modeling and simulation are among the most significant
developments in the practice of scientific inquiry in the 20th Century. Within the
last two decades, scientific computing has become an important contributor to
all scientific disciplines.
It is particularly important for the solution of research problems that are
insoluble by traditional scientific theoretical and experimental approaches,
hazardous to study in the laboratory, or time consuming or expensive to solve
by traditional means”
— “Scientific Discovery through Advanced Computing” DOE Office of
Science, 2000
3
Simulation: The Third Pillar of
Science
• Traditional scientific and engineering paradigm:
1)Do theory or paper design.
2) Perform experiments or build system.
• Limitations:
—Too difficult -- build large wind tunnels.
—Too expensive -- build a throw-away passenger jet.
—Too slow -- wait for climate or galactic evolution.
—Too dangerous -- weapons, drug design, climate experimentation.
• Computational science paradigm:
3) Use high performance computer systems to simulate the
phenomenon
– Base on known physical laws and efficient numerical methods.
4
Some Particularly Challenging
Computations
• Science
—Global climate modeling
—Biology: genomics; protein folding; drug design
—Astrophysical modeling
—Computational Chemistry
—Computational Material Sciences and Nanosciences
• Engineering
—Semiconductor design
—Earthquake and structural modeling
—Computation fluid dynamics (airplane design)
—Combustion (engine design)
—Crash simulation
• Business
—Financial and economic modeling
—Transaction processing, web services and search engines
• Defense
—Nuclear weapons -- test by simulations
—Cryptography
5
Technology Trends: Microprocessor
Capacity
2X transistors/Chip Every 1.5 years
Called “Moore’s Law”
Microprocessors have
become smaller, denser,
and more powerful.
Gordon Moore (co-founder of
Intel) predicted in 1965 that
the transistor density of
semiconductor chips would
double roughly every 18
months.
Slide source: Jack Dongarra
6
More Limits: How fast can
a serial computer be?
1 Tflop/s, 1
Tbyte sequential
machine

Consider the 1 Tflop/s sequential machine:



Data must travel some distance, r, to get from memory to
CPU.
To get 1 data element per cycle, this means 1012times per
second at the speed of light, c = 3x108 m/s.
Thus r <
c/1012 = 0.3 mm.
Now put 1 Tbyte of storage in a 0.3 mm x 0.3 mm
area:


r = 0.3
mm
Each bit occupies about 1 square Angstrom, or the
size of a small atom.
No choice but parallelism
7
Why Parallelism is now necessary for
Mainstream Computing
• Chip density is continuing
increase
~2x every 2 years
—Clock speed is not
—Number of processor
cores have to double
instead
• There is little or no
hidden parallelism (ILP) to
be found
• Parallelism must be
exposed to and
managed by software
Source: Intel, Microsoft (Sutter) and
Stanford (Olukotun, Hammond)
8
Fundamental limits on Serial
Computing: Three “Walls”
• Power Wall
—Increasingly, microprocessor performance is limited by achievable power
dissipation rather than by the number of available integrated-circuit resources
(transistors and wires). Thus, the only way to significantly increase the
performance of microprocessors is to improve power efficiency at about the
same rate as the performance increase.
• Frequency Wall
—Conventional processors require increasingly deeper instruction pipelines to
achieve higher operating frequencies. This technique has reached a point of
diminishing returns, and even negative returns if power is taken into account.
• Memory Wall
—On multi-gigahertz symmetric processors --- even those with integrated
memory controllers --- latency to DRAM memory is currently approaching 1,000
cycles. As a result, program performance is dominated by the activity of moving
data between main storage (the effective-address space that includes main
memory) and the processor.
9
What is Parallel computing?


Parallel computing involves performing
parallel tasks using more than one computer.
Example in real life with related principles -book shelving in a library



Single worker
P workers with each worker stacking n/p books,
but with arbitration problem(many workers try to
stack the next book in the same shelf.)
P workers with each worker stacking n/p books,
but without arbitration problem (each worker work
on a different set of shelves)
10
Important Issues in parallel
computing

Task/Program Partitioning.


Data Partitioning.


How to split a single task among the processors so
that each processor performs the same amount of
work, and all processors work collectively to
complete the task.
How to split the data evenly among the processors
in such a way that processor interaction is
minimized.
Communication/Arbitration.

How we allow communication among different
processors and how we arbitrate communication
related conflicts.
11
Challenges
1.
2.
3.
4.
Design of parallel computers so that we
resolve the above issues.
Design, analysis and evaluation of parallel
algorithms run on these machines.
Portability and scalability issues related to
parallel programs and algorithms
Tools and libraries used in such systems.
12
Units of Measure in HPC
• High Performance Computing (HPC) units are:
—Flop: floating point operation
—Flops/s: floating point operations per second
—Bytes: size of data (a double precision floating point number is 8)
• Typical sizes are millions, billions, trillions…
Mega Mflop/s = 106 flop/sec Mbyte = 220 = 1048576 ~ 106 bytes
Giga Gflop/s = 109 flop/sec Gbyte = 230 ~ 109 bytes
Tera Tflop/s = 1012 flop/sec Tbyte = 240 ~ 1012 bytes
Peta Pflop/s = 1015 flop/sec Pbyte = 250 ~ 1015 bytes
Exa Eflop/s = 1018 flop/sec Ebyte = 260 ~ 1018 bytes
Zetta Zflop/s = 1021 flop/sec Zbyte = 270 ~ 1021 bytes
Yotta Yflop/s = 1024 flop/sec Ybyte = 280 ~ 1024 bytes
• See www.top500.org for current list of fastest machines
13
What is a parallel computer?




A parallel computer is a collection of processors that
cooperatively solve computationally intensive
problems faster than other computers.
Parallel algorithms allow the efficient programming of
parallel computers.
This way the waste of computational resources can
be avoided.
Parallel computer v.s. Supercomputer


supercomputer refers to a general-purpose computer that
can solve computational intensive problems faster than
traditional computers.
A supercomputer may or may not be a parallel computer.
14
Parallel Computers: Past and Present



1980’s Cray supercomputer was 20-100 times faster
than other computers(main frames, minicomputers)
in use. (The price of supercomputer is 10 times other
computers – worth it)
1990’s “Cray”-like CPU is 2-4 times as fast as a
microprocessor. (The price of supercomputer is 10-20
times a microcomputer – make no sense)
The solution to the need for computational power is a
massively parallel computers, where tens to
hundreds of commercial off-the-shelf processors are
used to build a machine whose performance is much
greater than that of a single processor.
15
Scale of Today’s HPC Systems
Manufacturer
Computer
Rmax(Gfl
ops)
Installation site
Country
Year
#Core
1 Fujitsu
K computer, SPARC64 VIIIfx
2.0GHz, Tofu interconnect
8162000
RIKEN Advanced Institute
for Computational Science
(AICS)
Japan
2011
548352
2 NUDT
NUDT TH MPP, X5670 2.93Ghz
6C, NVIDIA GPU, FT-1000 8C
2566000
National Supercomputing
Center in Tianjin
China
2010
186368
3 Cray Inc.
Jaguar (Cray XT5-HE Opteron Six
Core 2.6 GHz)
1.759e+0
6
Oak Ridge National
Laboratory
USA
2009
224162
4 Dawning
Dawning TC3600 Blade, Intel
X5650, NVidia Tesla C2050 GPU
1271000
National Supercomputing
Centre in Shenzhen (NSCS)
China
2010
120640
16
CSI’s High Performance Center
Neptune. (neptune.csi.cuny.edu)
– a gateway or interface system for CUNY users that are not within local area network at
the College of Staten Island.
– As a single, two socket, 2 x 4 = 8 core head-like node, Neptune's 8 Intel Clovertown
cores run at 3.16 GHz. Neptune has a total of 16 Gbytes of memory or 2 Gbytes per
core.
– Neptune is not generally to be used for numerically intensive calculation, but as a
secure jumping-off point to access the larger cluster systems described below.
– Neptune can also be used as an access point to submit jobs using some applications
(MATLAB for instance) to the batch schedulers on the others systems.
– It can also be used to run a number of serial applications for which a GUI is required or
convenient.
Athena (athena.csi.cuny.edu)
– 97 node Dell PowerEdge Cluster (1 headnode and 96 compute nodes)
• 1 Gbit ethernet internal network
– 96 Compute nodes (PowerEdge 1850)
• Two Intel Xeon dual processor chips operating at 2.8 GHz
• 8 Gbytes of memory
– 1 Head node (PowerEdge 2850)
• Two Intel Xeon dual processor chips operating at 2.8 GHz
• 4 Gbytes of memory
17
CSI’s High Performance Center
• Zeus
•
•
– supporting users running Gaussian03, and now also, the development of CPU-GPU applications
–11 node Dell PowerEdge Cluster
• 1 Gbit ethernet internal network
– 10 Compute nodes (PowerEdge 1850)
– Compute nodes 0-7
• two sockets with Intel 2.66 GHz quad-core Harpertown processors
• providing a total of eight cores per node
• 8 Harpertown nodes have 2 Gbytes of memory per core for a total of 16 Gbytes per node
• Each Harpertown node also has a 1 TByte disk drive (/state/partition1) for storing Gaussian
scratch files.
– Compute nodes 8-9
• two sockets with Intel 2.27 GHz Woodcrest dual-core processors
• a total of 6 Gbytes of memory
• each attached to their own NVIDIA Tesla S1070, 1U, 4-way GPU array via dual PCI-Express
2.0 cables to support integrated CPU-GPU computing.
• Each GPU (4 per 1U Tesla node) has 240, 32-bit floating-pointing units with a peak
performance of 1 teraflop (there are 30 64-bit units).
• Each GPU also has 4 Gbytes of GPU-local memory
– 1 Head node (PowerEdge 1850)
• 2 x 4 cores running at 1.86 GHz
18
CSI’s High Performance Center
Bob
•
– named in honor of Dr. Robert E. Kahn, an alumnus of the City College of New York who, along with
Vinton G. Cerf, invented the TCP/IP protocol
– a Dell PowerEdge system consisting of one head node and thirty compute nodes
• both a standard 1 Gbit Ethernet interconnect and a low-latency, Infiniband SDR
(10 Gbit/second) interconnect
– 30 Compute nodes
• the same type providing a total of 30 x 8 = 240 cores.
• Each compute node has 16 Gbytes of memory or 2 Gbytes of memory per core
– 1 Head node (PowerEdge 1850)
• two sockets of AMD Shanghai native quad-core processors running at 2.3 GHz
Andy
•
– named in honor of Dr. Andrew S. Grove, an alumnus of the City College of New York and one of
the founders of Intel
– an SGI ICE system consisting of several head and service nodes, and 45 dual-socket, compute
nodes
• The interconnect network is a dual DDR Infiniband (20 Gbit/second) network in which one
rail is used for storage and the other for processor communication
– 30 Compute nodes
• each with Intel 2.93 GHz quad-core Intel Core 7 (Nehalem) processors providing a total of
360 compute cores
• Each compute node has 24 Gbytes of memory or 3 Gbytes of memory per core
19
– has a Lustre parallel file system with 24 Tbytes of useable storage
CFP2006 Performance numbers
for various CUNY HPC Systems
20
Applications of Parallel Computing









Astrophysics(explore the evoluation of galaxies, analysis of
extremely large datasets from telescope).
Material sciences (eg superconductivity).
Biology, biochemistry, gene sequencing.
Medicine and human organ modeling (eg. to study the effects
and dynamics of a heart attack, developing new drugs and
cures for diseases).
Global weather prediction.
Visualization (eg movie industry, 3D animation).
Data Mining (optimizing business and marketing decisions).
Computational-Fluid Dynamics (CFD) for aircraft and automotive
vehicle design.
Computer security, cryptography
21
Global Climate Modeling Problem

Problem is to compute:
f(latitude, longitude, elevation, time) ->
temperature, pressure, humidity, wind velocity

Approach:
—Discretize the domain, e.g., a measurement point every 10 km
—Devise an algorithm to predict weather at time t+δt given t

Uses:
- Predict major events,
e.g., El Nino
- Use in setting air
emissions standards
Source: http://www.epm.ornl.gov/chammp/chammp.html
22
Global Climate Modeling Computation
• One piece is modeling the fluid flow in the atmosphere
—Solve Navier-Stokes equations
– Roughly 100 Flops per grid point with 1 minute timestep
• Computational requirements:
—To match real-time, need 5 x 1011 flops in 60 seconds = 8
Gflop/s
—Weather prediction (7 days in 24 hours) -> 56 Gflop/s
—Climate prediction (50 years in 30 days) -> 4.8 Tflop/s
—To use in policy negotiations (50 years in 12 hours) -> 288
Tflop/s
• To double the grid resolution, computation is 8x to 16x
• State of the art models require integration of atmosphere,
ocean, sea-ice, land models, plus possibly carbon cycle,
geochemistry and more
• Current models are coarser than this
23
What is a parallel algorithm?

A parallel algorithm is an algorithm designed
for a parallel computer.
24
Questions when combining
processor power






How does one combine processors efficiently?
Do processors work independently?
Do they cooperate? If they cooperate how do
they interact with each other?
How are the processors interconnected?
How can we make programs portable?
How does one program such machines so
that programs run efficiently and do not
waster resourses?
25
End of lecture 1
Thank you!
26