09 ParallelizationRecap PDF

09_ParallelizationRecap
October 9, 2018
1 09 Parallelization Recap
In [1]: from IPython.display import Image
import re
import numpy as np
import matplotlib.pyplot as plt
In [2]: %alias clean rm -f *.c *.exe *.py *.pyc *.s .*f95 *.o *.fo *log *.h *.cc *.mod *.ppm *.p
In [3]: %clean
In [4]: !cat $HOME/Dropbox/Slackware/gcc_vars.sh
export PATH=/usr/local/gcc-7.3.0/bin:$PATH
export COMPILER_PATH=/usr/local/gcc-7.3.0/bin
export LIBRARY_PATH=/usr/local/gcc-7.3.0/lib:/usr/local/gcc-7.3.0/lib64:$LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/local/gcc-7.3.0/lib:/usr/local/gcc-7.3.0/lib64:$LD_LIBRARY_PATH
export CPATH=/usr/local/gcc-7.3.0/include
export C_INCLUDE_PATH=/usr/local/gcc-7.3.0/include
export CPLUS_INCLUDE_PATH=/usr/local/gcc-7.3.0/include
export OBJC_INCLUDE_PATH=/usr/local/gcc-7.3.0/include
export CC=/usr/local/gcc-7.3.0/bin/gcc
export CXX=/usr/local/gcc-7.3.0/bin/g++
export FC=/usr/local/gcc-7.3.0/bin/gfortran
export CPP=/usr/local/gcc-7.3.0/bin/cpp
export MANPATH=/usr/local/gcc-7.3.0/share/man:${MANPATH}
1.1 Goals
1. To have a brief review of what we have discussed
2. To have a llok at few additional topics
as always, with an hands on approach
1
1.2 Outline
Section ??
Section ??
Section ??
Section ??
2 0 - Recap on CPU architecture, performance and parallel program-

ming
2.0.1 0.1 Architecture and performance constrains
Progress and performance of computer architecture have been dominated for decades by a
logscale growth of the number of transistors on a single socket:
In [5]: Image("pictures/1200px-Moore's_Law_Transistor_Count_1971-2016.png")
Out[5]:
The fundamental constrains on CPU performance are power and heat:
In [6]: Image("pictures/power2.png")
2
Out[6]:
which needs to be dissipated.

That spawned the multi core and many core paradigms which incorporated parallelism into
desktop computers:
In [7]: Image("pictures/more_cores.png")
Out[7]:
3
hence to get full advantage of current CPU you need to know a little bit of parallelism
CPUs are based on a relatively simple paradigm,
In [8]: Image("pictures/von_neumann.png")
Out[8]:
4
but a single core is a complex object:
In [9]: Image("pictures/onecore.png")
Out[9]:
5
Instructions and data must be continuously fed to the control and arithmetic units, so that
the speed of the memory interface poses a limitation on compute performance (von Neumann
bottleneck).
The architecture is inherently sequential, processing a single instruction with (possibly) a sin-
gle operand or a group of operands from memory.
And we have several layers of optimization:
In [10]: Image("pictures/perfo_dim.png")
Out[10]:
6
• Sharing of compute load with a communication protocol:
– (Gigabit) Ethernet
– High-speed fabric (e. g. Section ?? or Section ??).
– Network topology may be rather complex.
– Each node may be itself a NUMA archicture with separated multicore sockets.
In [11]: Image("pictures/networks.png")
Out[11]:
7
2.0.2 Memory layout
Computer memory is organized in a hierarchy of decreasing price and speed and increasing ca-
pacity. Caches are small and fast and main memory is big and slow.
So chaches need to map memory several memory locations (as a juggler does with balls).
Cache divided into cache lines: when memory is read into cache, a whole cache line is always
read at the same time.
Good if we have data locality: nearby memory accesses will be fast.
Typical cache line size: 64-128 bytes
• L1 (approx. 10-100 kB) 1-5 cycles

• L2 (256 kB- ) 10-20 cycles
• L3 (> 2 MB), slower
Programs often re-use nearby memory locations: - Temporal locality: probably we will need
again soon data from a location - Spatial locality: nearby addresses likely to be needed soon
Example: reuse nearby elements in dense matrix multiplication (one of the past examples)
How cheap are FLOPs?
Rough estimate for one socket of the Section ??
Intel(R) Xeon(R) Gold 6132 CPU @ 2.60GHz with avx512:
• 14 cores
• 8 vector lanes per core
• 14 cores * 2.6 GHz * 2 FMA * 32 IPC * 8 vector lanes ~ 1165 GFLOP/s
• 1165 GFLOP/s * 8 bytes ~ 9.3 TFLOP/s
8
Maximum RAM bandwidth~ 0.12 TB/s. Hence 9.3/0.12 ~ 77.5 FLOPS per memory access.
Therefore, on this platform to have a compute bound application you must achieve approximately
78 FLOPS x access.
Strip mining and Cache Blocking
Suppose we have the following code:
for (int i = 0; i < m; i++)

for (int j = 0; j < n; j++)
compute(a[i], b[j]);
at each step we use a b[j] then throw it away and load the next. For n large with respect to
cache lines size, performance is determined by memory bandwidth.
now suppose we partially unroll manually the inner loop:
for (int i = 0; i < m; i++)

for (int jj = 0; jj < n; jj += TILE)
for (int j = jj; j < jj + TILE; j++)
where TILE is the cache line size. This is called strip mining (too much jargon) and, among
other things may help the compiler to vectorize code. But we have still a memory bound snippet.
what happens if the loop are permuted?
for (int jj = 0; jj < n; jj += TILE)

for (int i = 0; i < m; i++)
for (int j = jj; j < jj + TILE; j++)
we load TILE elements from b[j] and then use it for all computations, thereby increasing
temporal locality. This is called also cache blocking with block of size TILE.
In [12]: Image("pictures/cache_block.png")
Out[12]:
9
Compiler optimization
Once we have set up a decent piece of code we can ask the copiler to help a little bit:
In [13]: !gcc -O1 -Q --help=optimizers | grep enabled | grep inline
-finline [enabled]
-finline-atomics [enabled]
-finline-functions-called-once [enabled]
broadly speaking, compiler optimization can:
• reduce the number of operations

• reduce their mathematical complexity
• increase the data locality of operations
• use architecture specific optimizationand instructions sets
Arrays of Structures vs Structures of arrays Suppose you to write a code simulating the the
behaviour of many particle system, under some physical model, that is integrating its equation of
motion. You need data structures to hold basic information such positions, velocities, etc.
One way could be:
typedef struct Particle

{
float position[3];
float velocitiy[3];
float force[3];
float mass;
char tag[20];
} particle
then would be easy to form arrays of particles particle System[nparticles] and operate on
them. Adding new features to particles would involve changing the structure and some functions
for specific tasks, with (possibly) little modifications.
Using C++, classes and OOP would make it even more portable.
However what happens when run into a loop like that:
for(int i = 0; i < nparticles; i++)

for(int j = 0; j < 3; j++)
System[i].position[j] += timestep * System[i].velocity[j]; // Euler int.
locality is not guaranteed and is very difficult for the compiler to optimise aggresively (and
prevent vectorization).
Alternatively:
10
#define maxp 1e6;
typedef ParticleSystem
{
float Positions = malloc(maxp * 3 * sizeof(float);
...
} System;
system mysystem System, *sysptr;

float *X = sysprt->X; // deference then access member: (*ptr_to_s).member
float *V = systptr->V;
...
for(int i = 0; i < nparticles; i++)

for(int j = 0; j < 3; j++)
X[i][j] += timestep * V[i][j];
this is a common wisdom for performance oriented development and is called to create struc-
tures of arrays instead of arrays fo structures (Section ??).
2.0.3 0.2 Parallel programming.

Section ??
Parallel Computers are (or used to be) classified according to categories first proposed by M.
Flynn in the 60s (see https://ieeexplore.ieee.org/document/5009071?reload=true). The classifica-
tion is done permutating on the presence of single or multiple instructions streams and data streams:
• SISD: Single Instruction, Single Data.
• SIMD: Single Instruction, Multiple Data.
• MISD: Multiple Instruction, Single Data.
• MIMD: Multiple Instruction, Multiple Data.
Parallel programming jargon

Band
amount of data moved over time unit. Measured in bytes s−1 (hard disks) and bit s−1 (sockets and
nodes)
Latency minimum time needed to assign a requested resource
Performance number of 64 bit floating point operations per second
Concurrency A condition of a system in which multiple tasks are logically active at one time.
Parallelism A condition of a system in which multiple tasks are actually active at one time."
Parallel application
An application composed of tasks that actually execute concurrently in order to (1) consider larger
problems in fixed time or (2) complete in less time for a fixed size problem.
11
Load imbalance
The slowest process determines when everyone is done. Time waiting for other processes to finish
is time wasted.
Communication overhead
A cost only incurred by the parallel program. Grows with the number of processes for collective
communication.
In [14]: Image("pictures/parallel.png")
Out[14]:
Amdhal Law
Tser Tser 1−α
Tpar = = Tser ∗ f s + f p ∗ = (α + ) ∗ Tser
p p p
Tser 1
S= =
Tpar α + 1−p α
Weak Scaling
S( P) → S( P, N )
Tser 1
S( P, N ) = =
Tpar α + 1−p α
N→∞⇒α→0
S( P, N )α→0 = P
12
3 1 - Vectorization
• Perform one operation on multiple elements of a vector
• Chunk-wise processing instead of element wise
• Can lower computational cost
• A form of Single Instruction Multiple Data (SIMD) parallelism

basically, in addition to adding cores, I also add more FPU units per core
In [15]: Image("pictures/inst_set.png")
Out[15]:
• 1997 with MMX instruction set: 80-bit registers.
• 2000+ SSE instruction sets (several versions of them, from SSE1 to SEE4.2): 128-bit registers.
• 2011, Intel Sandy Bridge, AVX instruction set: 256-bit registers.
• 2016, Intel AVX-512: 512-bit registers.

See also Section ??.
Vectorization requires to sue specific registers (XMM, YMM) and assembly or compiler intrisics
and switches.
On Intel, SSE/AVX intrinsic functions use the following naming convention:
13
_<vector_size/instruction set>_<intrin_op>_<prec_suffix>
• is mm for 128 bit vectors (SSE), mm256 for 256 bit vectors (AVX and AVX2), and mm512 for
AVX512.
• Declares the operation of the intrinsic function. I.e. add, sub, mul ...
• Indicates the datatype. ps is for float, pd for double, and ep is for integer datatypes: epi32
for signed 32 bit integer, epu16 for unsigned 16 bit integer ...
For instance a vector intrisic multiplication for double vectors with AVX:
_mm256_mul_pd(&a , &a)
Single precision square root:
_mm256_sqrt_pf(&a).
There are notable features not available with intrisics, including:
1. Random numbers
2. Integer division
Available by including immintrin.h.

for a complete list, see #here
GCC supports multiple architectures and different header files are provided to use intrisics:
• arm_neon.h - ARM Neon types & intrinsics

• altivec.h - Freescale Altivec types & intrinsics
• mmintrin.h - X86 MMX
• xmmintrin.h - X86 SSE1
• emmintrin.h - X86 SSE2
Example of data load with intrisics:
In [16]: %%writefile access_128.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
#include <x86intrin.h>
// g. mancini july 18
int main()
{
float x[4] = {10., 20., 30., 40.};
__m128 One = {10., 20., 30., 40.};
__m128 Two = _mm_load_ps(x);
__m128 Three = _mm_set_ps(10., 20., 30., 40.);
for(int v=0; v < 4; v++)
printf("%d %f %f %f %f\n",v,x[v],One[v],Two[v],Three[v]);
printf("--------------\n\n");
float y[8] = {10., 20., 30., 40., 50. ,60., 70., 80.};
14
__m128 Four = _mm_load_ps(y);
for(int i=0; i < 2; i++)
for(int v=0; v < 4; v++)
printf("%d %d %f %f \n",v,i*4+v,y[i*4+v],Four[i*4+v]);
return 0;
}
Writing access_128.c
In [17]: %%bash
source $HOME/Dropbox/Slackware/gcc_vars.sh
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -o access.exe access_128.c
./access.exe
rm -f access.exe
0 10.000000 10.000000 10.000000 40.000000

1 20.000000 20.000000 20.000000 30.000000
2 30.000000 30.000000 30.000000 20.000000
3 40.000000 40.000000 40.000000 10.000000
--------------
0 0 10.000000 10.000000
1 1 20.000000 20.000000
2 2 30.000000 30.000000
3 3 40.000000 40.000000
0 4 50.000000 10.000000
1 5 60.000000 20.000000
2 6 70.000000 30.000000
3 7 80.000000 40.000000
OpenMP SIMD extensions

From version 4.0 onwards it can handled by OpenMP pragmas:
#pragma omp simd <something>

!$OMP simd <something>
Intrisics vs OpenMP pragmas:
In [18]: %%writefile intr_vs_pragma.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
//g. mancini july 18
int main()
{
15
const int steps = 1e7;
double *arr = calloc(sizeof(double), steps);
printf("non vectorized\n");
double t0 = omp_get_wtime();
for(int i = 0; i < steps; i++)
arr[i] = (double ) i;
double time = 1000.*(omp_get_wtime() - t0);
printf("time=%f\n", time);
printf("intrisics\n");
int dim = 2; //SSE
t0 = omp_get_wtime();
for(int i = 0; i < steps/dim; i += dim)
{
__m128d tmp = {(double ) i+1,(double ) i};
_mm_storeu_pd(&arr[i], tmp);
}
if( steps % dim != 0)
for(int i = (steps/dim)*dim; i < steps; i++)
arr[i] = (double ) i;
time = 1000.*(omp_get_wtime() - t0);
printf("OpenMP\n");
#pragma omp simd
for(int i=0; i < steps; i++)
arr[i] = (double ) i;;
time = 1000.*(omp_get_wtime() - t0);
return 0;
}
Writing intr_vs_pragma.c
In [19]: %%bash
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -fopenmp -o intr_vs_pragma.exe intr_
echo $OMP_NUM_THREADS
./intr_vs_pragma.exe
rm -f intr_vs_pragma.exe
non vectorized
time=21.300893
16
intrisics
time=4.626742
OpenMP
time=9.601765
3.0.1 Exercise 1.0

1. Write a program for calculating π using intrisics. (Last time we see this example, I promise).
Compare with non vectorized.
In [20]: Image("pictures/gears.png")
Out[20]:
** Solution 1.0 **
In [21]: %%writefile integral_intrisic.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
int main()
{
const long steps = 1E8;
const double start = 0.;
const double end = 1.;
double I = 0.;
const double dx = (end-start) / ((double ) steps);
for(int i=0; i < steps; i++)
{
double mid_x = start + dx * ((double ) i + 0.5);
double mid_y = 1.0/(1. + mid_x*mid_x);
I += mid_y;
}
I = 4. * I * dx;
double time = 1000.* (omp_get_wtime() - t0);
printf("PI=%f, evaluated in %f ms\n",I, time);
17
I = 0.;
printf("Evalute PI in vectorized loop\n");
__m128d num = _mm_set_pd(4.,4.);
__m128d IV = _mm_set_pd(0.,0.);
for(int i = 0; i < steps; i = i+2)
{
__m128d x = _mm_set_pd(i*dx, (i+1)*dx);
IV =_mm_add_pd(IV,_mm_div_pd(num,1+_mm_mul_pd(x,x)));
}
for(int i = 0; i < 2; i++)
I += IV[i];
I = I * dx;
time = 1000.* (omp_get_wtime() - t0);
printf("PI=%f, evaluated in %f ms\n",I, time);
return 0;
}
Writing integral_intrisic.c
In [22]: %%bash
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -fopenmp -o integral_intrisic.exe in
./integral_intrisic.exe
PI=3.141593, evaluated in 690.445218 ms

Evalute PI in vectorized loop
PI=3.141593, evaluated in 345.213376 ms
4 2 - OpenMP
Shared memory computer : any computer composed of multiple processing elements that share
an address space.
Two Classes: - Symmetric multiprocessor (SMP): a shared address space with “equal-time”
access for each processor, and the OS treats every processor the same way. (Real SMP have not
existed for many years, see slides of class#1)
• Non Uniform address space multiprocessor (NUMA;you have one of these): different mem-
ory regions have different access costs ... think of memory segmented into “Near” and “Far”
memory.
Process - An instance of a program execution. - The execution context of a running program ...
i.e. the resources associated with a program’s execution.
In [23]: Image("pictures/process.png")
18
Out[23]:
Thread:
• Threads are "light weight processes”

• Threads share process state among multiple threads ... this greatly reduces the cost of switch-
ing context.
In [24]: Image("pictures/Thread.png")
Out[24]:
19
** A shared memory program**
• One process and lots of threads.

• Threads interact through reads/writes to a shared address space.
• An OS scheduler decides when to run which threads ... interleaved for fairness.
• Synchronization to assure every legal order results in correct results.
How do threads interact?
• Threads communicate by sharing variables.

• Unintended sharing of data causes race conditions:
– race condition: when the program’s outcome changes as the threads are scheduled
differently.
• To control race conditions:
– Use synchronization to protect data conflicts.
• Synchronization is expensive so:
20
– Change how data is accessed to minimize the need for synchronization.
OpenMP i. e. Open Multi Processing an Application Program Interface (API) for developing
multithreaded applications.
OpenMP core syntax
Most of the constructs in OpenMP are compiler directives. #pragma omp construct [clause
[clause]...]
Example #pragma omp parallel num_threads(4)
Function prototypes and types in the file: #include
Most OpenMP constructs apply to a structured block: - Structured block: a block of one or more
statements with one point of entry at the top and one point of exit at the bottom.
Parallel regions are created with OpenMP whenever you invoke #pragma omp parallel. From
that point until the end of the region (structured block closed or !$OMP end parallel) a team of
threads is spawned and the code is executed indepently by each thread.
In [25]: Image("pictures/fork_join.jpeg")
Out[25]:
• Unintended sharing of data causes race conditions in which the program’s results depend
on thread scheduling.
• To control race conditions: use synchronization to protect data conflicts.
• Synchronization is expensive so: change how data is accessed to minimize the need for
synchronization.
• Synchronization may use barriers (all threads in a team) or mutual exclusions (mutexes) (a
pair of threads)
Work sharing constructs
• Loop construct
• Single construct
• Master construct
• Sections construct
• Task construct
Sections construct
21
#pragma omp parallel
{
#pragma omp sections
{
#pragma omp section
{
// structured block 1
}
#pragma omp section
{
}
#pragma omp section
{
}
...
}
}
sections starts the construct. It contains several section constructs which marks a different
block, containing a represents a task. (Beware of section and sections).
sections distributes the blocks/tasks between existing threads. The requirement is that each
block must be independent of the other blocks. Then each thread executes one block at a time.
Each block is executed only once by one thread.
Environment Variables
• OMP_NUM_THREADS number of threads
• OMP_SCHEDULE (when using schedule(runtime) in loop worksharing constructs)
• OMP_THREADLIMIT maximum number of threads
• OMP_STACKSIZE maximum thread stack size
• GOMP_CPU_AFFINITY: how threads are bound to CPUs; e. g. GOMP_CPU_AFFINITY="0

2 1 3"
• OMP_PROC_BIND: allows (or not) to move threads; overrides affinity
• OMP_DISPLAY_ENV
see also the Section ??
Example: π integration with reduction (I lied)
In [26]: %%writefile integral_omp_red.cc

#include <iostream>
22
#include <omp.h>
int main(int argc, char **argv)
{
const int num_steps = 1e8;

double I = 0.;
const double dx = (end-start) / ((double ) num_steps);
#pragma omp parallel for reduction(+:I) schedule(static,1)

for(int i=0; i < num_steps; i++)
{
double mid_y = 4.0/(1. + mid_x*mid_x) ;
I += mid_y;
}
I = I * dx;
std::cout << "Loop completed in "<< 1000.*(t1-t0) << " ms using " << std::endl;
std::cout << "Integral value = " << I << std::endl;
}
Writing integral_omp_red.cc
In [27]: %%bash
g++ -Wall -O2 -fopenmp -march=native -mtune=native -o integral.exe integral_omp_red.cc
OMP_NUM_THREADS=1 ./integral.exe
Loop completed in 690.51 ms using
Integral value = 3.14159
In [28]: %%capture pibench

%%bash
export LD_LIBRARY_PATH=/usr/local/gcc-7.3.0/lib64/:$LD_LIBRARY_PATH
In [29]: pat = re.compile(r'(Loop\scompleted\sin\s)(.+)\sms')
result = pat.findall(pibench.stdout)
result = np.array(list(map(lambda x:float(x[-1]),result)))
23
In [30]: plt.plot((1,2,4,8),result,marker='s',color='k',ls='-')
plt.ylabel('Wall time')
plt.xlabel('number of threads')
Out[30]: Text(0.5,0,'number of threads')
4.0.1 Exercise 2.0

Mandelbrot set
The Mandelbrot set is an iterative map in the complex plane:
z2n+1 = z2n + c
where c is a constant and z0 = 0.
Points that do not diverge after a finite number of iterations are part of the set.
1. Write a serial code for the Mandelbrot set.

2. Write an OpenMP version.
In [31]: %%writefile template.c

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <omp.h>
void main(int argc, char *argv[])
{
//settings
const int maxiter = 500;
double horiz = 2.0;
const int width = 1024;
const int height = 1024;
const double xmin = -2.;
const double xmax = 1.;
const double ymin = -1.2;
const double ymax = 1.2;
const char* filename = "out.ppm";
//variables
int k;
horiz = horiz*horiz;
double x0, y0, xres, yres;
xres = (xmax - xmin)/((double) width);
yres = (ymax - ymin)/((double) height);
double re, im, re2, im2;
// open output PPM file, https://en.wikipedia.org/wiki/Netpbm_format#PPM_example

unsigned char colour[3];
24
FILE * fp = fopen(filename,"w");
fprintf(fp,"P3\n%d %d %d\n",width, height, 255);
//compute map for every pixel
//write black & white out file
}
Writing template.c
Out[32]:
** Solution 2.0 **
In [33]: %%writefile mandelbrot_serial.c

// g mancini apr 18
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <string.h>
#include <omp.h>
inline int mandel(double x0, double y0, double thres, int maxiter)
{
double re = 0.0;
double im = 0.0;
double re2 = 0.0;
double im2 = 0.0;
int k;
for (k = 0; k < maxiter && (re2 + im2 < thres); k++)
{
im = 2. * re * im + y0;
re = re2 - im2 + x0;
re2 = re * re;
im2 = im * im;
}
return k;
}
void write_output(char* filename, int* image, int width, int height, int maxiter)
25
{

for (int i=0; i< height; i++)

{
for (int j=0; j < width; j++)
{
if(image[i + j*height] == maxiter)
fprintf(fp,"%d %d %d ",0,0,0);
else
fprintf(fp,"%d %d %d ",255,255,255);
}
fprintf(fp,"\n");
}
fclose(fp);
}
int main(int argc, char *argv[])

{
//settings
double thres = 3.0;
char* filename = "out.ppm";
//variables
thres = thres*thres;
double x0, y0, xres, yres;
int *image = malloc(sizeof(int) * height * width);
//calculate
y0 = ymin;
for (int i=0; i< height; i++)
{
x0 = xmin;
{
image[i + j*height] = mandel(x0,y0,thres,maxiter);
x0 = x0 + xres;
26
}
y0 = y0 + yres;
}// close on i
printf("Spent %f seconds\n",omp_get_wtime()-t0);
write_output( filename, image, width, height, maxiter);
free(image);
}
Writing mandelbrot_serial.c
In [34]: %%bash
rm -f out.*
gcc -Wall -O2 -march=native -mtune=native -fopenmp -o mandelbrot.exe mandelbrot_serial.
./mandelbrot.exe
#https://www.imagemagick.org/script/index.php
convert out.ppm out.png
Spent 0.290289 seconds
In [35]: Image("out.png")
Out[35]:
27
In [36]: %%timeit
%%bash
./mandelbrot.exe

1 loop, best of 3: 468 ms per loop
In [37]: !rm -f mandelbrot.exe
28
OpenMP version
In [38]: %%writefile mandelbrot_omp.c

// g mancini apr 18
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <omp.h>
inline int mandel(double x0, double y0, double tr, int maxiter)
{
double re = 0.0;
double im = 0.0;
double re2 = 0.0;
double im2 = 0.0;
int k;
for (k = 1; k < maxiter && (re2 + im2 < tr); k++)

{
im = 2. * re * im + y0;
re = re2 - im2 + x0;
re2 = re * re;
im2 = im * im;
}
return k;
}
void write_output(char* filename, int* image, int width, int height, int maxiter)
{
for (int i = 0; i < height; i++)

{
{
if(image[i + j*height] >= maxiter)
fprintf(fp,"%d %d %d ",0,0,0);
else
fprintf(fp,"%d %d %d ",255,255,255);
}
fprintf(fp,"\n");
}
fclose(fp);
}
29
{
//settings
char* filename = "out.ppm";
double horiz = 3.0;
//variables
horiz = horiz*horiz;
double xres, yres;
//allocate image
int *image = malloc(sizeof(int) * height * width);
// run
#ifdef _OPENMP
int block = 8;
#pragma omp parallel for schedule(dynamic) //collapse(2)
for(int jj = 0; jj < width; jj+=block)
for (int i = 0; i < height; i++)
{
for (int j = jj; j < jj + block; j++)
{
double x0 = xmin + j*xres;
double y0 = ymin + i*yres;
image[i + j*height] = mandel(x0,y0,horiz,maxiter);
}
}// close on i
printf("spent %f ms in parallel loop\n",(t1-t0)*1000.0);
#else
double y0 = ymin;
for (int i = 0; i< height; i++)
{
double x0 = xmin;
for (int j = 0; j < width; j++)
{
image[i + j*height] = mandel(x0,y0,horiz,maxiter);
x0 = x0 + xres;
30
}
y0 = y0 + yres;
}// close on i
#endif
write_output( filename, image, width, height, maxiter);
free(image);
return 0;
}
Writing mandelbrot_omp.c
In [39]: %%bash
rm -f out.*
gcc -Wall -O2 -march=native -mtune=native -fopenmp -o mandelbrot.exe mandelbrot_omp.c -
OMP_NUM_THREADS=2 ./mandelbrot.exe
convert out.ppm out.png
spent 951.497681 ms in parallel loop
In [40]: Image("out.png")
Out[40]:
31
In [41]: %%capture thr1
%%bash
In [42]: thr1.stdout
Out[42]: 'spent 1902.723774 ms in parallel loop\n'
In [43]: %%bash
32
In [44]: %%bash
In [45]: %%capture thr8

%%bash
In [46]: thr8.stdout
Out[46]: 'spent 289.876555 ms in parallel loop\n'
Final Speedup
In [47]: t1 = float(thr1.stdout.split()[1])
t8 = float(thr8.stdout.split()[1])
print("Final speedup ",t1/t8)
Final speedup 6.563910537711475
4.0.2 OpenMP tasks

What will be printed by the following snippet?

{
{
printf("A ");
printf("race ");
printf("car ");
}
printf("\n");
return 0;
}
output depends on number of threads and execution sequence. The word order will be mixed.

{
{
#pragma omp single
{
printf("A ");
printf("race ");
printf("car ");
33
}
printf("\n");
return 0;
}
this solves the problem but all threads but one are wasted. Using the task construct:
In [48]: %%writefile sanrossore.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
int main()
{
{
#pragma omp single
{
printf("A ");
#pragma omp task
printf("race ");
#pragma omp task
printf("horse ");
} //single thread
} //parallel region
printf("\n");
return 0;
}
Writing sanrossore.c
In [49]: %%bash
gcc -Wall -O2 -fopenmp -o race.exe sanrossore.c
OMP_NUM_THREADS=1 ./race.exe
A race horse
In [50]: %%bash
A race horse
Tasks are independent units of work. Threads are assigned to perform the work of each task.
Tasks are composed of:
- code to execute
34
- data environment (it own its data)
- internal control variables
Task barrier (taskwait): encountering thread suspends until all child tasks it has generated are
completed.
Tasks can be executed in arbitrary order and generated until there is some work to do:
In [51]: Image("pictures/tasks.png")
Out[51]:
taskwait sets up some final work to do:
In [52]: %%writefile sanrossore.c

#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
int main()
{
{
#pragma omp single
{
printf("A ");
35
#pragma omp task
printf("race ");
#pragma omp task
printf("car ");
#pragma omp taskwait
printf(" or a horse one?");
} //single thread
} //parallel region
printf("\n");
return 0;
}
Overwriting sanrossore.c
In [53]: %%bash
A race car or a horse one?
4.0.3 (Parallel) Design Patterns

Design Patterns - Qualitative description of a frequently occurring problem in some domain and
its solution - (object oriented programming, network apps...)
Description usually involves problem definition, driving forces, solution, benefits, difficulties,
related patterns.
Pattern Language - a collection of design patterns, guiding the users through the decision
process in building a system.
• Parallel is hard and seldom exploits full hardware capabilities.

• A “cookbook” illustrating common problems and solutions would be handy (loop centric,
divide and conquer ...)
• Provide a common vocabulary (factory pattern).
• Improve software reusability.
Flowchart:
• FindingConcurrency
– data vs. control
• AlgorithmStructure
– pipeline, replicate...
• SupportingStructure
– SPMD, fork/join...
36
• Implementation
– barriers, locks...
Common patterns:
• Embarassingly Parallel (Mandelbrot set)

• Replicable (Integral)
• Repository
• Divide and Conquer (Fibonacci with recursion)
• Pipeline
• Recursive Data
• Geometric
• IrregularMesh
• Inseparable
The Divide and Conquer design pattern

Divide and conquer is an important design pattern with two distinct phases
• Divide phase:
– Breaks down problem into two or more sub-problems of the same (or related) type.
– Continue division until these sub-problems become simple enough to be solved

directly.
• Conquer phase:
– Executes the computations on each of the “indivisible” sub-problems.
– May also combine solutions of sub-problems until the solution of the original problem
is reached.
Implementation is typically done with recursive algorithms. Why?
• Because the nature of recursion forms smaller sub-problems that are very much like the
larger problem being solved.
• The return from recursive calls can be used to combine partial solutions into an overall so-
lution.
How to implement it?

- Define a split operation - Continue to split the problem until subproblems are small enough to
solve directly (you fix a threshold).
- Recombine solutions to subproblems to solve original global problem.
the Fibonacci sequence with tasks is an example of the Divide and Conquer design pattern: we
spit until a known value is returned then merge again.
In [54]: Image("pictures/dac.jpg")
Out[54]:
37
Fibonacci sequence with tasks
int fib ( int n ) // n is private between tasks when called recursively

{
int x,y;
if ( n < 2 ) return n;
#pragma omp task shared (x)
x = fib(n-1);
#pragma omp task shared(y)
y = fib(n-2);
#pragma omp taskwait // explicit task barrier
return x+y; // if x and y weren't shared they would stay with the inizialita
}
The QuickSort algorithm Section ?? is a Divide and Conquer algorithm.

It picks an element as pivot and partitions an input array around it, forming two new subarrays. DaC
is applied by calling iteratively QuickSort on the subarrays up to arrays of lenght=2.
The critical part is the partition procedure which moves the elements wrt to the pivot. Given
an array A and a pivot p, partition, elements greater than the pivot go to the right and elements
smaller (or equal) go to the left.
Different implementations of QuickSort pick the pivot in different ways; e. g.:
1. Pick first element.
2. Pick last element.
3. Pick a random element.
4. Pick median (my choice but not for any insight).
In [55]: Image("pictures/Quicksort.png")
Out[55]:
38
In [56]: %%writefile quicksort.c
#include <stdlib.h>
#include <stdio.h>
//https://rosettacode.org/wiki/Sorting_algorithms/Quicksort#C
void quicksort(int *A, int len)

{
if (len < 2) return;
int i, j, temp;
//median element
int pivot = A[len / 2];
for (i = 0, j = len - 1; ; i++, j--)
{
//update position of last element to be moved on the right
while (A[i] < pivot) i++;
//update position of last element to be moved on the left
while (A[j] > pivot) j--;
if (i >= j) break;
temp = A[i];
A[i] = A[j];
A[j] = temp;
39
}
//left branch
quicksort(A, i);
//right branch
quicksort(A + i, len - i);
}
void printA(int *A, int L)

{
for (int i = 0; i < L; i++)
printf("%d ", A[i]);
printf("\n");
}
int main (int argc, char *argv[])

{
int A[] = {10, 7, 8, 9, 1, 5, 7, 4};
int length = sizeof(A)/sizeof(A[0]);
printf("Sorted array:\n");
printA(A, length);
printf("Unsorted array:\n");
quicksort(A, length);
printA(A, length);
return 0;
}
Writing quicksort.c
In [57]: %%bash
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -o QS.exe quicksort.c
./QS.exe
Sorted array:
10 7 8 9 1 5 7 4
Unsorted array:
1 4 5 7 7 8 9 10
4.0.4 Exercise 2.1

Write an OpenMP version of QuickSort using tasks.
once you created a working version, try to use bigger arrays; fill it with random integers using
the Section ?? random number generators; see Section ??:
gsl_rng *R = gsl_rng_alloc (gsl_rng_taus);

gsl_rng_set(R, 211085);
40
printf("Generating %d random integers in[%ld,%ld]\n",length,gsl_rng_min(R),gsl_rng_max(R));
for( int i = 0; i < length; i++)
Arr[i] = gsl_rng_uniform_int(R, length);
gsl_rng_free(R)
compile adding linker and preprocessor flags, e. g.:
gcc ... -lm -lgsl -lgslcblas -I/usr/include/gsl/
add timing and some super simple check.
Out[58]:
4.0.5 Solution 3.1

In [59]: %%writefile quicksort_omp.c
#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
#include <gsl_rng.h>
//https://rosettacode.org/wiki/Sorting_algorithms/Quicksort#C
void quicksort(int *A, int len)

{
if (len < 2) return;
int i, j, temp;
//median element
int pivot = A[len / 2];
for (i = 0, j = len - 1; ; i++, j--)
{
//update position of last element to be moved on the right
while (A[i] < pivot) i++;
//update position of last element to be moved on the left
while (A[j] > pivot) j--;
if (i >= j) break;
temp = A[i];
41
A[i] = A[j];
A[j] = temp;
}
//left branch
#pragma omp task
quicksort(A, i);
//right branch
#pragma omp task
quicksort(A + i, len - i);
}
void printA(int *A, int L)

{
for (int i = 0; i < L; i++)
printf("%d ", A[i]);
printf("\n");
}
int main (int argc, char *argv[])

{
int length = atoi(argv[1]);
int *A = malloc(length * sizeof(int));
gsl_rng *R = gsl_rng_alloc (gsl_rng_taus);
gsl_rng_set(R, 211085);
printf("Generating %d random integers in[%ld,%d]\n",length,gsl_rng_min(R),length);
for( int i = 0; i < length; i++)
A[i] = gsl_rng_uniform_int(R, length);
gsl_rng_free(R);
#ifdef VRB
printf("Unsorted array:\n");
printA(A, length);
#endif
int nt;
{
nt = omp_get_num_threads();
{
#pragma omp single
quicksort(A, length);
}
}
printf("Sorted in %f with %d threads\n",omp_get_wtime()-t0,nt);

#ifdef VRB
42
printf("Sorted array:\n");
printA(A, length);
#endif
return 0;
}
Writing quicksort_omp.c
In [60]: %%bash
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -o QS.exe quicksort_omp.c \
-fopenmp -lm -lgsl -lgslcblas -I/usr/include/gsl/ -DVRB
OMP_NUM_THREADS=1 ./QS.exe 20
Generating 20 random integers in[0,20]

Unsorted array:
5 17 1 5 5 13 4 10 15 13 15 8 19 2 19 5 6 19 3 13
Sorted in 0.000022 with 1 threads
Sorted array:
1 2 3 4 5 5 5 5 6 8 10 13 13 13 15 15 17 19 19 19
In [61]: %%bash

Unsorted array:
5 17 1 5 5 13 4 10 15 13 15 8 19 2 19 5 6 19 3 13
Sorted array:
1 2 3 4 5 5 5 5 6 8 10 13 13 13 15 15 17 19 19 19
In [62]: %%bash
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -o QS.exe quicksort_omp.c \
-fopenmp -lm -lgsl -lgslcblas -I/usr/include/gsl/
In [63]: %%bash

In [64]: %%bash
43
In [65]: %%bash

In [66]: %%bash

5 3 - MPI
MPI (Message Passing Interface) is standard specification for message passing.
• Implemented in several libraries, open source and proprietary.

– Section ?? and Section ?? commonly used (and available in distro repositories ...)
• Designed for distributed computing (but it can be used for shared multiprocessing).
• Portable.
• C, C++ and Fortran bindings available.
• Currently, version 3 of the standard available.
MPI features
• The Program consists of a collection of processes (often counted by their rank).

• Number of processes almost always fixed at program startup time.
• Each node has a local address space.
• Logically shared data must be partioned.
• Processes communicate by explicit send/receive pairs.
When an MPI program is run, multiple processes are executed and work of separate blocks of
data.
The collection of processes involved in a computationis called a process group.
Data is shared either with one-to-one:
44
• MPI_Send(),
• MPI_Recv()
or collective communications:
• MPI_Barrier()
• MPI_Gather()
• MPI_Scatter()
Actually, there are a lot of functions:
In [67]: Image("pictures/mpi_funcs.png")
Out[67]:
Example: Jacobi Solver
In [68]: %%writefile jacobi.c

#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <mpi.h>
#include <string.h>
#include <omp.h>
double check_jacobi(int dim, double **A, double *b, double *xnew)
45
{
double sum = 0.;
for(int i = 0; i < dim; i++)
{
double sum_i = b[i];
for(int j = 0; j < dim; j++)
sum_i -= A[i][j] * xnew[j];
sum += pow(sum_i,2);
}
return sqrt(sum/(double ) dim);
}
void gen_jac_input(int DIM, int high, double **A, double *b, double *xo)
{
for (int i=0; i < DIM;i++)
{
b[i] = (double ) (rand() % high);
xo[i] = (double ) (rand() % high);
double sum = 0.;
for (int j=0; j< DIM; j++)

{
A[i][j] = (double ) (rand() % high);
sum += A[i][j];
}
A[i][i] = 1.5 * (sum - A[i][i]);
}
}
double do_iter(int rank, int first, int last, int dim, \

double **A, double *b, double *xold, double *xnew)
{
double conv = 0.;
for(int i = first; i < last; i++)
{
double tmp = 0.;
int row = i + rank*last;
#pragma omp simd reduction(+:tmp)
for(int j = 0; j < dim; j++)
tmp += xold[j] * A[row][j];
xnew[i] = (b[row] - tmp + xold[row] * A[row][row])/A[row][row];
conv += pow(xnew[i]-xold[row],2);
}
return conv;
}
double jacobi(int dim, double **A, double *b, double *xold,\

double *xnew, double tol, int kmax, int *niter)
46
{
int rank, size;

#ifdef MPI_VERSION
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
#else
rank = 0;
size = 1;
#endif
double gconv = 1. + tol;

int chunk = dim / size;
int k = 0;
while(k < kmax && gconv > tol)

{
double conv = do_iter(rank, 0, chunk, dim, A, b, xold, xnew);
#ifdef MPI_VERSION
if(rank == 0 && dim % size > 0)
{
conv += do_iter(0, size*chunk, dim, dim, A, b, xold, xnew);
for(int i = size*chunk; i < dim; i++)
xold[i] = xnew[i];
}
MPI_Allgather(xnew, chunk, MPI_DOUBLE, xold, chunk, MPI_DOUBLE,
MPI_COMM_WORLD);
if(dim % size > 0)

MPI_Bcast(&xold[size*chunk], dim % size, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Allreduce(&conv, &gconv, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
#else
gconv = conv;
#endif
k++;
}
memcpy(xnew,xold,sizeof(double) * dim);
*niter = k;
return gconv;
}

{
double tol = 1e-8;
const int kmax = 100;
const int DIM = atoi(argv[1]);
const int high = atoi(argv[2]);
int niter;
47
double t0 = 0., t1 = 0.;
srand(23);
//allocate
double **A = malloc(DIM * sizeof(double*));
A[0] = malloc(DIM * DIM * sizeof(double));
for(int i = 1; i < DIM; i++)
A[i] = A[0] + i*DIM;
double *b = malloc(DIM * sizeof(double));
double *xo = malloc(DIM * sizeof(double));
double *xn = calloc(DIM, sizeof(double));
//generate A and b and make A diagonally dominant

gen_jac_input(DIM, high, A, b, xo);
int rank = 0;
MPI_Init(&argc, &argv);
//call jacobi
t0 = MPI_Wtime();
double conv = jacobi(DIM, A, b, xo, xn, tol, kmax, &niter);
MPI_Finalize();
t1 = MPI_Wtime();
if(rank == 0)
{
if(niter < kmax)
printf("converged at %g after %d iterations in %g s\n",conv,niter,t1-t0);
else
printf("NOT converged (conv=%g) after %d iterations\n",conv,niter);
//check results
double error = check_jacobi(DIM, A, b, xn);
printf("error=%f\n",error);
}
}
Writing jacobi.c
In [69]: %%bash
mpicc -O2 -Wall -march=native -mtune=native -msse4.2 -lm -fopenmp -lmpi -o jacobi.exe j
mpirun -n 1 jacobi.exe 4000 100
48
converged at 8.85065e-09 after 45 iterations in 0.559049 s
error=0.294429
In [70]: %%bash

error=0.294429
In [71]: %%bash

error=0.294429
5.0.1 Non blocking communications

One can improve performance on many systems by overlapping communication and computation.
A nonblocking send starts initiates the send operation, but does not complete it, returning before the
message copy is completed.
A nonblocking receive initiates the receive operation but returns storing the complete message.
Separates communication into three steps: 1. Initiate non-blocking communication
- returns immediately - routine name starting with MPI_I
2. Do some work ( - Local to the processes) - “latency hiding”
3. Wait for non-blocking communication to complete
int MPI_Isend(void *buffer, int count, MPI_Datatype datatype, int dest,

int tag, MPI_Communicator comm, MPI_Request *request);
int MPI_Irecv(void *buffer, int count, MPI_Datatype datatype, int source, int tag,
MPI_Communicator comm, MPI_Request *request);
MPI_Request takes the role of MPI_Status in the blocking Recv but now is needed also by the
send call.
Completition of the request is made with MPI_Test:
int MPI_Test(MPI_Request *request, int *flag, MPI_Status *status);
ISend/Test is then equivalent to Send.

testing can be delayed with MPI_Wait():
int MPI_Wait(MPI_Request *request, MPI_Status *status);
Example: non blocking ring
49
In [72]: %%writefile ring_nb.c
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <mpi.h>
{
int rank, size, check, to, from;
const char hdr[11] = ": sent to ";

const char ftr[20] = " and received from ";
char msg[33], rmsg[33] = " xxx" ;
MPI_Request ReqS, ReqR;

MPI_Status Stat;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD,&rank);
MPI_Comm_size(MPI_COMM_WORLD,&size);
check = rank % size;

if (check == 0)
{
to = 1;
from = size -1;
}
else if (check == size-1)
{
to = 0;
from = rank -1;
}
else
{
to = check + 1;
from = check - 1;
}
snprintf(msg,sizeof(msg),"%d%s%d%s%d",rank,hdr,to,ftr,from);
if (size > 1)
{
MPI_Isend(&msg[0], 35, MPI_BYTE, to, check, MPI_COMM_WORLD, &ReqS);
MPI_Irecv(rmsg,35, MPI_BYTE, from, from, MPI_COMM_WORLD, &ReqR);
MPI_Wait(&ReqS, &Stat);
MPI_Wait(&ReqR, &Stat);
printf("%s \n",rmsg);
}
MPI_Finalize();
50
}
Writing ring_nb.c
In [73]: %%bash
mpicc -O2 -Wall -o ring.exe ring_nb.c -march=native -mtune=native -lm -lmpi
mpirun -n 8 ring.exe
7: sent to 0 and received from 6

Geometric decomposition pattern

Dependencies exist, but communicate in predictable (geometric) neighbour-to-neighbour paths.
Use Neighbour to Neighbour communications.
Examples: Finite differences, CFD, Molecular Dynamics.
In [74]: Image("pictures/geom_decomp.png")
Out[74]:
51
Example: 1D heat equation
In [75]: Image("pictures/rod.png")
Out[75]:
52
Use first order and second order finite difference for time and space, respectively:
∂T T n +1 − T n
=
∂t δt
∂ T
2 Tj+1 − 2. · Tj + Tj−1
=
∂x 2 δx
then combine:
Tjn+1 = Tjn − r · ( Tjn+1 − 2. · Tjn + Tjn−1 )

where
δt
r=
δx2
must be less than 0.5 to converge.
How to parallelise this?
Answer: Geometric Decomposition pattern: Break it into chunks assigning one chunk to each process.
Problem: end of chunk cells depend on each other across processes.
Solution: use ghost cells to replicate data between processes.
In [76]: %%writefile heat.c

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <mpi.h>
//gmancini may 18
int main()
{
double H0 = 100., H1 = 100.;
int Ncell = 5e2;

int Nstep = 1e6;
double delta_t, delta_x, scalef;

double *Temp, *Tnew, *swap;
int rank, size, chunk, first, last;
MPI_Init(0,NULL);
MPI_Status status;
if (size > 1)
{
if(Ncell % size != 0)
Ncell = Ncell + Ncell % size;
53
chunk = Ncell / size + 2;
first = 2;
last = chunk-2;
Temp = calloc(chunk, sizeof(double));
Tnew = calloc(chunk, sizeof(double));
if(rank == 0)
Temp[1] = H0;
if(rank == size-1)
Temp[chunk-2] = H1;
}
else
{
first = 1;
last = Ncell - 1;
chunk = Ncell;
Temp = calloc(Ncell, sizeof(double));

Tnew = calloc(Ncell, sizeof(double));
Temp[0] = H0;
Temp[Ncell-1] = H1;
}
for(int i = first-1; i <= last+1; i++)

Tnew[i] = Temp[i];
delta_x = 1. / ((double ) Ncell);

delta_t = 1. / ((double ) Nstep);
scalef = delta_t / (delta_x * delta_x);
if(rank == 0)
printf("scalef, Ncell, chunk: %f %d %d \n",scalef,Ncell,chunk);
for(int time = 0; time < Nstep; time++)

{
if (size > 1)
{
// has a left neighbour
if(rank != 0)
{
MPI_Send(&Temp[1], 1, MPI_DOUBLE, rank-1, rank, MPI_COMM_WORLD);
MPI_Recv(&Temp[0], 1, MPI_DOUBLE, rank-1, rank-1, MPI_COMM_WORLD, &stat
}
// has a right neighbour
if(rank != size - 1)
{
MPI_Recv(&Temp[last+1], 1, MPI_DOUBLE, rank+1, rank+1, MPI_COMM_WORLD,
MPI_Send(&Temp[last], 1, MPI_DOUBLE, rank+1, rank, MPI_COMM_WORLD);
54
}
} //endif size > 1
//apply stencil
for(int cell = first; cell < last; cell++)
Tnew[cell] = Temp[cell] + scalef * (Temp[cell+1] - 2.*Temp[cell] + Temp[cel
if(size > 1)
{
// 0 is a ghost cell
if(rank != 0)
Tnew[1] = Temp[1] + scalef * (Temp[2] - 2.*Temp[1] + Temp[0]);
// chunk+2-1 is a ghost cell

Tnew[last] = Temp[last] + scalef * (Temp[last+1] - 2.*Temp[last] + Temp
}
swap = Tnew;
Tnew = Temp;
Temp = swap;
} //end for time
if (size > 1)
{
swap = calloc(Ncell, sizeof(double));
MPI_Allgather(&Tnew[1],last,MPI_DOUBLE,swap,last,MPI_DOUBLE,MPI_COMM_WORLD);
}
MPI_Finalize();
if (rank == 0)
{
char outname[30] = "temp_field_np_";
char SZ[10];
sprintf(SZ,"%d",size);
strcat(outname, SZ);
FILE *fp = fopen(outname, "w");

for(int i = 0; i < Ncell; i++)
fprintf(fp,"%d %f\n", i, swap[i]);
fclose(fp);
}
return 0;
}
Writing heat.c
In [77]: %%bash
55
mpicc --version
mpicc -Wall -O2 -march=native -mtune=native -lm -lmpi -o heat.exe heat.c
gcc (GCC) 7.3.0

Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
In [78]: %%bash
mpirun -n 2 ./heat.exe
scalef, Ncell, chunk: 0.250000 500 252
In [79]: %%bash
mpirun -n 4 ./heat.exe
In [80]: %%bash
paste temp_field_np_2 temp_field_np_4 | awk 'NR>1{a+=($2-$4)*($2-$4)}END{print a/(NR-1)
5.0.2 Exercise 3.0

Write a new version of heat.c using non blocking communications.
Out[81]:
56
5.0.3 Solution 3.0
In [82]: %%writefile heat_nb.c
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <mpi.h>
//g mancini may 18
int main()
{
double H0 = 100., H1 = 100.;
int Ncell = 5e2;

int Nstep = 1e6;
double delta_t, delta_x, scalef;

double *Temp, *Tnew, *swap;
int rank, size, chunk, first, last;
MPI_Init(0,NULL);
MPI_Request Requests[4];
if (size > 1)
{
if(Ncell % size != 0)
Ncell = Ncell + Ncell % size;
chunk = Ncell / size + 2;
first = 2;
last = chunk-2;
Temp = calloc(chunk, sizeof(double));
Tnew = calloc(chunk, sizeof(double));
if(rank == 0)
Temp[1] = H0;
if(rank == size-1)
Temp[chunk-2] = H1;
}
else
{
first = 1;
last = Ncell - 1;
chunk = Ncell;
Temp = calloc(Ncell, sizeof(double));

Tnew = calloc(Ncell, sizeof(double));
57
Temp[0] = H0;
Temp[Ncell-1] = H1;
}
for(int i = first-1; i <= last+1; i++)

Tnew[i] = Temp[i];
delta_x = 1. / ((double ) Ncell);

delta_t = 1. / ((double ) Nstep);
scalef = delta_t / (delta_x * delta_x);
if(rank == 0)
printf("scalef, Ncell, chunk: %f %d %d \n",scalef,Ncell,chunk);
for(int time = 0; time < Nstep; time++)

{
//Send data
if (size > 1)
{
//has a left neighbour
if(rank != 0)
MPI_Irecv(&Temp[0], 1, MPI_DOUBLE, rank-1, rank-1,
MPI_COMM_WORLD, &Requests[0]);
//has a right neighbour
MPI_Irecv(&Temp[last+1], 1, MPI_DOUBLE, rank+1, rank+1,
//has a left neighbour
if(rank != 0)
MPI_Isend(&Temp[1], 1, MPI_DOUBLE, rank-1, rank,
//has a right neighbour
MPI_Isend(&Temp[last], 1, MPI_DOUBLE, rank+1, rank,
} //endif size > 1
//apply stencil
for(int cell = first; cell < last; cell++)
Tnew[cell] = Temp[cell] + scalef * (Temp[cell+1] - 2.*Temp[cell] + Temp[cel
//receive
if(size > 1)
{
if(rank != 0)
{
MPI_Wait(&(Requests[0]),MPI_STATUS_IGNORE);
58
Tnew[1] = Temp[1] + scalef * (Temp[2] - 2.*Temp[1] + Temp[0]);
}
{
Tnew[last] = Temp[last] + scalef * (Temp[last+1]
- 2.*Temp[last] + Temp[last-1]);
}
}
swap = Tnew;
Tnew = Temp;
Temp = swap;
} //end for time
if (size > 1)
{
swap = calloc(Ncell, sizeof(double));
MPI_Allgather(&Tnew[1],last,MPI_DOUBLE,swap,last,MPI_DOUBLE,MPI_COMM_WORLD);
}
MPI_Finalize();
if (rank == 0)
{
char outname[30] = "temp_field_np_";
char SZ[10];
sprintf(SZ,"%d",size);
strcat(outname, SZ);
FILE *fp = fopen(outname, "w");
for(int i = 0; i < Ncell; i++)
fprintf(fp,"%d %f\n", i, swap[i]);
fclose(fp);
}
return 0;
}
Writing heat_nb.c
In [83]: %%bash
mpicc --version
mpicc -Wall -O2 -march=native -mtune=native -lm -lmpi -o heat_nb.exe heat_nb.c
gcc (GCC) 7.3.0

Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
59
In [84]: %%bash
mpirun -n 4 ./heat_nb.exe
In [85]: %%bash
paste temp_field_np_2 temp_field_np_4 | awk 'NR>1{a+=($2-$4)}END{print a/(NR-1)}'
5.0.4 Combining MPI and OpenMP

Combining MPI and OpenMP in theory can give benefits:
1. use non local resources for even bigger problems

2. use light weight thread and locality in sockets with OpenMP
... yes, I lied
In [86]: %%writefile integral_mpi_omp.c

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <mpi.h>
#include <omp.h>
{
long num_steps = 1e8;
const double dx = (end-start) / ((double ) num_steps);
double PI = 0.;
int nt;
int rank, size, provided;

MPI_Init_thread(0,NULL,MPI_THREAD_FUNNELED,&provided);
if(rank==0)
printf("Provided mode %d\n",provided);
double t0 = MPI_Wtime();
if(num_steps % size != 0)
60
num_steps += num_steps/size;
int chunk = num_steps / size;
int i0 = rank*chunk;
int i1 = (rank+1)*chunk;
double I = 0.;
#pragma omp parallel for simd reduction(+:I)
for(int i = i0; i < i1; i++)
{
nt = omp_get_num_threads();
double mid_y = 4.0/(1. + mid_x*mid_x) ;
I += mid_y;
}
MPI_Reduce(&I, &PI, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
PI = PI * dx;
double t1 = MPI_Wtime();
MPI_Finalize();
if(rank==0)
{
printf("Loop completed in %f using %d procs and %d threads\n",1000.*(t1-t0),siz
printf("Integral value %f\n",PI);
}
}
Writing integral_mpi_omp.c
In [87]: %%bash
gcc -Wall -O2 -march=native -mtune=native -msse4.2 -fopenmp -fopenmp-simd -fopt-info-ve
-o I.exe integral_mpi_omp.c -lm -lmpi
integral_mpi_omp.c:35:38: note: loop vectorized
In [88]: %%bash
OMP_NUM_THREADS=4 mpirun -n 2 ./I.exe
Provided mode 1
Loop completed in 104.758742 using 2 procs and 4 threads
Integral value 3.141593
Caveats
1. Messages are sent to a process not to a particular thread.
61
Not all MPIs are threadsafe. MPI 2.0 defines threading modes:
- MPI_Thread_Single: no support for multiple threads.
- MPI_Thread_Funneled: Multiple threads, only master calls MPI.
- MPI_Thread_Serialized: Mult threads each calling MPI, but they do it one at a time.
- MPI_Thread_Multiple: Multiple threads without any restrictions.
Check out threading mode with:
MPI_init_thread(desired_mode, delivered_mode, ierr)
Environment variables are not propagated by mpirun.
Youll need to broadcast OpenMP parameters and set them with the library routines.
5.0.5 Exercise 3.1

Add multithreading to heat.c. Increase problem size and run some benchmarks.
Out[89]:
5.0.6 Solution 3.1

almost trivial
just add #pragma omp parallel for simd above the stencil.
From Tim Mattson:
Literature is mixed on the hybrid model: sometimes its better, sometimes MPI alone is best.
There is potential for benefit to the hybrid model because:
1. MPI algorithms often require replicated data making them less memory efficient.
2. Fewer total MPI communicating agents means fewer messages and less overhead from
message conflicts.
3. Algorithms with good cache efficiency should benefit from shared caches of multi-threaded
programs.
The model maps perfectly with clusters of SMP nodes. But really, its a case by case basis and
to large extent depends on the particular application.
62

09 ParallelizationRecap PDF

Uploaded by

Copyright:

Available Formats

09 ParallelizationRecap PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

09 ParallelizationRecap PDF

Uploaded by

Copyright:

Available Formats

09_ParallelizationRecap

In [4]: !cat $HOME/Dropbox/Slackware/gcc_vars.sh

as always, with an hands on approach

2 0 - Recap on CPU architecture, performance and parallel program-

The fundamental constrains on CPU performance are power and heat:

which needs to be dissipated.

– Network topology may be rather complex.

• L1 (approx. 10-100 kB) 1-5 cycles

for (int i = 0; i < m; i++)

for (int i = 0; i < m; i++)

for (int jj = 0; jj < n; jj += TILE)

In [13]: !gcc -O1 -Q --help=optimizers | grep enabled | grep inline

broadly speaking, compiler optimization can:

• reduce the number of operations

typedef struct Particle

for(int i = 0; i < nparticles; i++)

system mysystem System, *sysptr;

for(int i = 0; i < nparticles; i++)

2.0.3 0.2 Parallel programming.

• SISD: Single Instruction, Single Data.

• SIMD: Single Instruction, Multiple Data.

• MISD: Multiple Instruction, Single Data.

• MIMD: Multiple Instruction, Multiple Data.

Parallel programming jargon

• Chunk-wise processing instead of element wise

• Can lower computational cost

• A form of Single Instruction Multiple Data (SIMD) parallelism

• 1997 with MMX instruction set: 80-bit registers.

• 2011, Intel Sandy Bridge, AVX instruction set: 256-bit registers.

• 2016, Intel AVX-512: 512-bit registers.

Single precision square root:

There are notable features not available with intrisics, including:

Available by including immintrin.h.

• arm_neon.h - ARM Neon types & intrinsics

Example of data load with intrisics:

In [16]: %%writefile access_128.c

0 10.000000 10.000000 10.000000 40.000000

OpenMP SIMD extensions

#pragma omp simd <something>

Intrisics vs OpenMP pragmas:

In [18]: %%writefile intr_vs_pragma.c

3.0.1 Exercise 1.0

In [21]: %%writefile integral_intrisic.c

PI=3.141593, evaluated in 690.445218 ms

• Threads are "light weight processes”

• One process and lots of threads.

How do threads interact?

• Threads communicate by sharing variables.

Work sharing constructs

• OMP_NUM_THREADS number of threads

• OMP_SCHEDULE (when using schedule(runtime) in loop worksharing constructs)

• OMP_THREADLIMIT maximum number of threads

• OMP_STACKSIZE maximum thread stack size

• GOMP_CPU_AFFINITY: how threads are bound to CPUs; e. g. GOMP_CPU_AFFINITY="0

• OMP_PROC_BIND: allows (or not) to move threads; overrides affinity

see also the Section ??

Example: π integration with reduction (I lied)

In [26]: %%writefile integral_omp_red.cc

const int num_steps = 1e8;