Transcript Presentazione di PowerPoint

Corso di Laurea Magistrale e
Dottorati in Matematica Applicata
CALCOLO SCIENTIFICO
(PARALLELO)
Prof. Luca F. Pavarino
Dipartimento di Matematica
Universita` di Milano
a.a. 2009-2010
[email protected],
http://www.mat.unimi.it/~pavarino
1
Struttura del corso
• Orario
- Lunedi` 10.30-12.30
Aula 2
- Mercoledi` 8.30-11.30
- Giovedi` 13.30-14.30
Aula 2
Aula 9 (compattare?)
• 12-13 settimane (~ ½ lezione, ½ laboratorio)
software
hardware
• Laboratorio in Aula 2: esercitazioni con
-
Nostro Cluster Linux (ulisse.mat.unimi.it), 104 processori
Nostro nuovo Cluster Linux Nemo
Cluster Linux del Cilea (avogadro.cilea.it), ~1700 processori
(Cluster Linux del Cineca (bcx.cineca.it), 5120 processori)
- (nuovo IBM SP6 del Cineca (sp6.sp.cineca.it))
- Uso della libreria standard per “message passing” MPI
- Uso della libreria parallela di calcolo scientifico PETSc
dell’Argonne National Lab., basata su MPI
2
Materiale e Testi
•
Slides in inglese basate su corsi di calcolo parallelo tenuti a
Univ. Illinois da Michael Heath, UC Berkeley da Jim Demmel,
(+ MIT da Alan Edelmann)
•
Possibili testi:
- A. Grama, A. Gupta, G. Karipys, V. Kumar, Introduction to parallel
computing, 2nd ed., Addison Wesley, 2003
- L. R. Scott, T. Clark, B. Bagheri, Scientific Parallel Computing,
Princeton University Press, 2005
•
Molto materiale on-line, e.g.:
-
www-unix.mcs.anl.gov/dbpp/ (Ian Foster’s book)
www.cs.berkeley.edu/~demmel/ (Demmel’s course)
www-math.mit.edu/~edelman/ (Edelman’s course)
www.cse.uiuc.edu/~heath/ (Heath’s course)
www.cs.rit.edu/~ncs/parallel.html (Nan’s ref page)
3
Schedule of Topics
1. Introduction
2. Parallel architectures
3. Networks
4. Interprocessor communications: point-to-point, collective
5. Parallel algorithm design
6. Parallel programming, MPI: message passing interface
7. Parallel performance
8. Vector and matrix products
9. LU factorization
10. Cholesky factorization
11. PETSc parallel library
12. Iterative methods for linear systems
13. Nonlinear equations and ODEs
14. Partial Differential Equations
15. Domain Decomposition Methods
16. QR factorization
17. Eigenvalues
4
1) Introduction
• What is parallel computing
• Large important problems require powerful computers
• Why powerful computers must be parallel processors
• Why writing (fast) parallel programs is hard
• Principles of parallel computing performance
5
What is parallel computing
• It is an example of parallel processing:
- division of task (process) into smaller tasks (processes)
- assign smaller tasks to multiple processing units that work
simultaneously
- coordinate, control and monitor the units
• Many examples from nature:
- human brain consists of ~10^11 neurons
- complex living organisms consist of many cells (although monocellular
organism are estimated to be ½ of the earth biomass)
- leafs of trees ...
• Many examples from daily life:
-
highways tollbooths, supermarket cashiers, bank tellers, …
elections, races, competitions, …
building construction
written exams ...
6
• Parallel computing is the use of multiple processors to
execute different parts of the same program (task)
simultaneously
• Main goals of parallel computing are:
- Increase the size of problems that can be solved
- bigger problem would not be solvable on a serial computer in a
reasonable amount of time  decompose it into smaller problems
- bigger problem might not fit in the memory of a serial computer 
distribute it over the memory of many computer nodes
- Reduce the “wall-clock” time to solve a problem
 Solve (much) bigger problems (much) faster
Subgoal: save money using cheapest available
resources (clusters, beowulf, grid computing,...)
7
Not at all trivial that more processors help to achieve these
goals:
• “If a man can dig a hole of 1 m3 in 1 hour, can 60 men dig
the same hole in 1 minute (!) ? Can 3600 men do it in 1
second (!!) ?”
• “I know how to make 4 horses pull a cart, but I do not
know how to make 1024 chickens do it” (Enrico Clementi)
• “ What happens if the mean-time to failure for nodes on
the Tflops machine is shorter than the boot time ?
(Courtenay Vaughan)
8
Why we need
powerful computers
9
Simulation: The Third Pillar of Science
• Traditional scientific and engineering method:
(1) Do theory or paper design
(2) Perform experiments or build system
Theory
• 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
Experiment
Simulation
• Computational science and engineering paradigm:
(3) Use high performance computer systems
to simulate and analyze the phenomenon
- Based on known physical laws and efficient numerical methods
- Analyze simulation results with computational tools and
methods beyond what is used traditionally for experimental
data analysis
10
Some Particularly Challenging Computations
• Science
- Global climate modeling, weather forecasts
- Astrophysical modeling
- Biology: Genome analysis; protein folding (drug design)
- Medicine: cardiac modeling, physiology, neurosciences
• Engineering
-
Airplane design
Crash simulation
Semiconductor design
Earthquake and structural modeling
• Business
- Financial and economic modeling
- Transaction processing, web services and search engines
• Defense
- Nuclear weapons (ASCI), cryptography, …
11
$5B World Market in Technical Computing
1998 1999 2000 2001 2002 2003
100%
90%
80%
70%
Other
Technical Management and
Support
Simulation
Scientific Research and R&D
Mechanical
Design/Engineering Analysis
Mechanical Design and
Drafting
60%
Imaging
50%
Geoscience and Geoengineering
40%
Electrical Design/Engineering
Analysis
Economics/Financial
30%
Digital Content Creation and
Distribution
20%
Classified Defense
10%
Chemical Engineering
0%
Biosciences
Source: IDC 2004, from NRC Future of Supercomputer Report
12
Units of Measure in HPC
• High Performance Computing (HPC) units are:
- Flops: floating point operations
- 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
bytes
Giga
Tera
Peta
Exa
Zetta
Yotta
Mflop/s = 106 flop/sec
Mbyte = 220 = 1048576 ~ 106
Gflop/s = 109 flop/sec
Tflop/s = 1012 flop/sec
Pflop/s = 1015 flop/sec
Eflop/s = 1018 flop/sec
Zflop/s = 1021 flop/sec
Yflop/s = 1024 flop/sec
Gbyte = 230 ~ 109 bytes
Tbyte = 240 ~ 1012 bytes
Pbyte = 250 ~ 1015 bytes
Ebyte = 260 ~ 1018 bytes
Zbyte = 270 ~ 1021 bytes
Ybyte = 280 ~ 1024 bytes
Current fastest (public) machine ~ 1.5 Pflop/s
Up-to-date lisy at www.top500.org
13
Ex. 1: Global Climate Modeling Problem
•
Problem is to compute:
f(latitude, longitude, elevation, time) 
temperature, pressure, humidity, wind velocity
•
Atmospheric model: equation of fluid dynamics

Navier-Stokes system of nonlinear partial differential equations
•
Approach:
-
Discretize the domain, e.g., a measurement point every 1km
Devise an algorithm to predict weather at time t+1 given t
• Uses:
- Predict major events,
e.g., El Nino
- Use in setting air
emissions standards
14
Source: http://www.epm.ornl.gov/chammp/chammp.html
Global Climate Modeling Computation
• One piece is modeling the fluid flow in the atmosphere
- Solve Navier-Stokes problem
- Roughly 100 Flops per grid point with 1 minute timestep
• Computational requirements:
- To match real-time, need 5x 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 at least 8x
• 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
15
Climate Modeling on the Earth Simulator System
 Development of ES started in 1997 in order to make a
comprehensive understanding of global environmental
changes such as global warming.
 Its construction was completed at the end of February,
2002 and the practical operation started from March 1,
2002
 35.86Tflops (87.5% of the peak performance) is achieved in the
Linpack benchmark.
 26.58Tflops was obtained by a global atmospheric circulation
code.
16
Ex. 2: Cardiac simulation
• Very difficult problem spanning many disciplines:
- Electrophysiology (spreading of electrical excitation front)
- Structural Mechanics (large deformation of incompressible
biomaterial)
- Fluid Dynamics (flow of blood inside the heart)
• Large-scale simulations in computational
electrophysiology (joint work with P. Colli-Franzone)
- Bidomain model (system of 2 reaction-diffusion equations) coupled
with Luo-Rudy 1 gating (system of 7 ODEs) in 3D
- Q1 finite elements in space + adaptive semi-implicit method in time
- Parallel solver based on PETSc library
- Linear systems up to 36 M unknowns each time-step (128 procs of
Cineca SP4) solved in seconds or minutes
- Simulation of full heartbeat (4 M unknowns in space, thousands of
time-steps) took more than 6 days on 25 procs of Cilea HP
Superdome, then about 50 hours on 36 procs of our cluster, now 6.5
hours using multilevel preconditioner
17
3D simulations: isochrones of acti, repo, APD
18
• Hemodynamics in circulatory system (work in
Quarteroni’s group at MOX, Polimi)
• Blood flow in the heart (Peskin’s group, CIMS, NYU)
- Modeled as an elastic structure in an incompressible fluid.
- The “immersed boundary method” due to Peskin and McQueen.
- 20 years of development in model
- Many applications other than the heart: blood clotting, inner ear,
paper making, embryo growth, and others
- Use a regularly spaced mesh (set of points) for evaluating the fluid
- Uses
-
Current model can be used to design artificial heart valves
Can help in understand effects of disease (leaky valves)
Related projects look at the behavior of the heart during a heart attack
Ultimately: real-time clinical work
19
Ex. 3: latest breakthrough
20
21
22
23
24
25
26
Ex. 4: Parallel Computing in Data Analysis
• Web search:
-
Functional parallelism: crawling, indexing, sorting
Parallelism between queries: multiple users
Finding information amidst junk
Preprocessing of the web data set to help find information
• Google physical structure (2004 estimate, check
current status on e.g. wikipedia):
- about 63.272 nodes (126,544 cpus)
- 126.544 GB RAM
- 5,062 TB hard drive space
(This would make Google server farm one of the most powerful
supercomputer in the world)
• Google index size (June 2005 estimate):
- about 8 billion web pages, 1 billion images
27
- Note that the total Surface Web ( = publically indexable, i.e.
reachable by web crawlers) has been estimated (Jan. 2005) at
over 11.5 billion web pages.
- Invisible (or Deep) Web ( = not indexed by search engines; it
consists of dynamic web pages, subscription sites, searchable
databases) has been estimated (2001) at over 550 billion
documents.
- Invisible Web not to be confused with Dark Web consisting of
machines or network segments not connected to the Internet
• Data collected and stored at enormous speeds
(Gbyte/hour)
- remote sensor on a satellite
- telescope scanning the skies
- microarrays generating gene expression data
- scientific simulations generating terabytes of data
- NSA analysis of telecommunications
28
Why powerful
computers are
parallel
29
Tunnel Vision by Experts
• “I think there is a world market for maybe five
computers.”
- Thomas Watson, chairman of IBM, 1943.
• “There is no reason for any individual to have
a computer in their home”
- Ken Olson, president and founder of Digital Equipment
Corporation, 1977.
• “640K [of memory] ought to be enough for
anybody.”
- Bill Gates, chairman of Microsoft,1981.
Slide source: Warfield et al.
30
Technology Trends: Microprocessor Capacity
Moore’s Law
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
31
Impact of Device Shrinkage
• What happens when the feature size shrinks by a factor
of x ?
• Clock rate goes up by x
- actually less than x, because of power consumption
• Transistors per unit area goes up by x2
• Die size also tends to increase
- typically another factor of ~x
• Raw computing power of the chip goes up by ~ x4 !
- of which x3 is devoted either to parallelism or locality
32
Microprocessor Transistors per Chip
• Growth in transistors per chip
• Increase in clock rate
100,000,000
1000
10,000,000
1,000,000
i80386
i80286
100,000
R3000
R2000
100
Clock Rate (MHz)
Transistors
R10000
Pentium
10
1
i8086
10,000
i8080
i4004
1,000
1970 1975 1980 1985 1990 1995 2000 2005
Year
0.1
1970
1980
1990
2000
Year
33
But there are limiting forces
Manufacturing costs and yield problems limit use of density
•
Moore’s 2nd law
(Rock’s law): costs go
up
Demo of
0.06
micron
CMOS
Source: Forbes Magazine
•
Yield
-What percentage of the chips
are usable?
-E.g., Cell processor (PS3) is
sold with 7 out of 8 “on” to
improve yield
34
34
Revolution is Happening Now
• Chip density is
continuing
increase ~2x
every 2 years
- Clock speed is not
- Number of
processor cores
may double instead
• There is little or
no more hidden
parallelism (ILP)
to be found
• Parallelism must
be exposed to
and managed by
software
Source: Intel, Microsoft (Sutter) and
Stanford (Olukotun, Hammond)
35
Parallelism in 2009?
• These arguments are no longer theoretical
• All major processor vendors are producing multicore chips
- Every machine will soon be a parallel machine
- To keep doubling performance, parallelism must double
• Which commercial applications can use this parallelism?
- Do they have to be rewritten from scratch?
• Will all programmers have to be parallel programmers?
- New software model needed
- Try to hide complexity from most programmers – eventually
- In the meantime, need to understand it
• Computer industry betting on this big change, but does not
have all the answers
- Berkeley ParLab established to work on this
36
Physical limits: how fast can a serial computer be?
1 Tflop/s, 1 Tbyte
sequential
machine
r = 0.3 mm
• Consider the 1 Tflop/s sequential machine:
- Data must travel some distance, r, to get from memory to CPU.
- Go get 1 data element per cycle, this means 1012 times 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 0.3 mm area:
(in fact 0.3^2 mm^2/10^12 = 9 10^(-2) 10^(-6) m^2/10^12 =
9 10^(-20) m^2 = (3 10^(-10))^2 m^2 = 3^2 A^2 
- Each byte occupies less than 3 square Angstroms, or the size of a
small atom! (1 Angstrom = 10^(-10) m = 0.1 nanometer)
• No choice but parallelism
37
More Exotic Solutions on the Horizon
•
GPUs - Graphics Processing Units (eg NVidia)
- Parallel processor attached to main processor
- Originally special purpose, getting more general
•
FPGAs – Field Programmable Gate Arrays
- Inefficient use of chip area
- More efficient than multicore now, maybe not later
- Wire routing heuristics still troublesome
•
Dataflow and tiled processor architectures
- Have considerable experience with dataflow from 1980’s
- Are we ready to return to functional programming languages?
•
Cell
- Software controlled memory uses bandwidth efficiently
- Programming model not yet mature
38
“Automatic” Parallelism in Modern Machines
• Bit level parallelism: within floating point operations, etc.
• Instruction level parallelism (ILP): multiple instructions execute per
clock cycle.
• Memory system parallelism: overlap of memory operations with
computation.
• OS parallelism: multiple jobs run in parallel on commodity SMPs.
There are limitations to all of these:
 to achieve high performance, the programmer needs to identify,
schedule and coordinate parallel tasks and data.
39
Processor-DRAM Gap (latency)
CPU
“Moore’s Law”
Processor-Memory
Performance Gap:
(grows 50% / year)
DRAM
DRAM
7%/yr.
100
10
1
µProc
60%/yr.
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
Performance
1000
Time
40
Principles of Parallel Computing
• Parallelism and Amdahl’s Law
• Finding and exploiting granularity
• Preserving data locality
• Load balancing
• Coordination and synchronization
• Performance modeling
All of these issues makes parallel programming harder
than sequential programming.
41
Amdahl’s law: Finding Enough Parallelism
• Suppose only part of an application seems parallel
• Amdahl’s law
- Let s be the fraction of work done sequentially, so
(1-s) is fraction parallelizable.
- P = number of processors.
Speedup(P) = Time(1)/Time(P)
<= 1/(s + (1-s)/P)
<= 1/s
Even if the parallel part speeds up perfectly, we may be
limited by the sequential portion of code.
Ex: if only s = 1%, then speedup <= 100
 not worth it using more than p = 100 processors
42
Overhead of Parallelism
• Given enough parallel work, this is the most significant
barrier to getting desired speedup.
• Parallelism overheads include:
-
cost of starting a thread or process
cost of communicating shared data
cost of synchronizing
extra (redundant) computation
• Each of these can be in the range of milliseconds
(= millions of flops) on some systems
• Tradeoff: Algorithm needs sufficiently large units of work
to run fast in parallel (i.e. large granularity), but not so
large that there is not enough parallel work.
43
Locality and Parallelism
Conventional
Storage
Proc
Hierarchy
Cache
L2 Cache
Proc
Cache
L2 Cache
Proc
Cache
L2 Cache
L3 Cache
L3 Cache
Memory
Memory
Memory
potential
interconnects
L3 Cache
• Large memories are slow, fast memories are small.
• Storage hierarchies are large and fast on average.
• Parallel processors, collectively, have large, fast memories -- the slow accesses to
“remote” data we call “communication”.
• Algorithm should do most work on local data.
44
Load Imbalance
• Load imbalance is the time that some processors in the
system are idle due to
- insufficient parallelism (during that phase).
- unequal size tasks.
• Examples of the latter
- adapting to “interesting parts of a domain”.
- tree-structured computations.
- fundamentally unstructured problems
- Adaptive numerical methods in PDE (adaptivity and parallelism seem
to conflict).
• Algorithm needs to balance load
- but techniques to balance load often reduce locality
45
Measuring Performance: Real Performance?
Peak Performance grows exponentially,
a la Moore’s Law

In 1990’s, peak performance increased 100x; in
2000’s, it will increase 1000x
1,000
But efficiency (the performance relative to
the hardware peak) has declined

was 40-50% on the vector supercomputers of
1990s
now as little as 5-10% on parallel
supercomputers of today
Close the gap through ...


Mathematical methods and algorithms that
achieve high performance on a single
processor and scale to thousands of
processors
More efficient programming models and tools
for massively parallel supercomputers
100
Teraflops

Peak Performance
Performance
Gap
10
1
Real Performance
0.1
1996
2000
2004
46
Performance Levels
• Peak advertised performance (PAP)
- You can’t possibly compute faster than this speed
• LINPACK
- The “hello world” program for parallel computing
- Solve Ax=b using Gaussian Elimination, highly tuned
• Gordon Bell Prize winning applications performance
- The right application/algorithm/platform combination plus years of work
• Average sustained applications performance
- What one reasonable can expect for standard applications
When reporting performance results, these levels are
often confused, even in reviewed publications
47
Performance Levels (for example on NERSC-5)
• Peak advertised performance (PAP): 100 Tflop/s
• LINPACK (TPP): 84 Tflop/s
• Best climate application: 14 Tflop/s
- WRF code benchmarked in December 2007
• Average sustained applications performance: ? Tflop/s
- Probably less than 10% peak!
• We will study performance
- Hardware and software tools to measure it
- Identifying bottlenecks
- Practical performance tuning (Matlab demo)
48
49
50
51
52
53
54
55
Simple example 1: sum of N numbers, P procs
N
A   ai
i 1
Also known as reduction
(of the vector [a1,…,aN] to the scalar A)
- Assume N is an integer multiple of P: N = kP
- Divide the sum into P partial sums:
Aj 
jk
a
i
i ( j 1) k 1
P
Then
P parallel tasks, each with
k -1 additions of k = N/P data
A   Aj
j 1
Global sum (not parallel,
communication needed)
56
Simple example 2: pi
1
2
1
4
/(
1

x
)
dx

4
arctg
(
x
)
|
0 

0
N
- Use composite midpoints quadrature rule:
where h = 1/N and xi  (i  1 / 2)h
2
4
h
/(
1

x

i ),
i 1
-Decompose sum into P parallel partial
sums + 1 global sum, (as before or with
stride P)
On processor myid = 0,…,P-1, (P = numprocs) compute:
sum = 0;
for I = myid + 1:numprocs:N,
x = h*(I – 0.5);
sum = sum + 4/(1+x*x);
end;
mypi = h*sum;
global sum the local mypi into glob_pi (reduction)
57
Simple example 3: prime number sieve
See exercise in class
Simple example 4: Jacobi method for BVP
See exercise in class
58