Introduction to MPI
Wednesday, February 10, 16
Topics to be covered
• MPI vs shared memory
• Initializing MPI
• MPI concepts — communicators, processes,
ranks
• MPI functions to manipulate these
• Timing functions
• Barriers and the reduction collective operation
Wednesday, February 10, 16
Shared and distributed memory
• Shared memory
• automatically maintained a consistent image of memory according
to some memory model
• fine grained communication possible via loads, stores, and cache
coherence
• model and multicore hardware support well aligned
• Programs can be converted piece-wise
• Distributed memory
• Program executes as a collection of processes, all communication
between processors explicitly specified by the programmer
• Fine grained communication in general too expensive —
programmer must aggregate communication
• Conversion of programs is all-or-nothing
• Cost scaling of machines is better than with shared memory — well
aligned with economics of commodity rack mounted blades
Wednesday, February 10, 16
Message Passing
network –
ethernet or
proprietary (vendor
specific, infinitband,
etc.)
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
Wednesday, February 10, 16
Message Passing Model
network
– ethernet or
proprietary (vendor
specific, infinitband,
etc.)
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
processor
memory
• This drawing implies
that all processor are
equidistant from one
another
• This is often not the
case — the network
topology and
multicores make some
processors closer
than others
• programmers have to
exploit this manually
Wednesday, February 10, 16
Message Passing Model
• In reality, processes run on
cores, and are closer to
other processes on the
same processor
• Across processors, some
can be reached via a single
hop on the network, others
require multiple hops
• Not a big issue on small
(several hundred
processors), but it needs to
be considered on large
machines.
network
P
M
P
M
P
M
P
M
network
P
M
P
M
P
M
P
M
network
P
M
P
M
P
M
P
M
network
P
M
P
M
P
M
P
M
network
Wednesday, February 10, 16
131,072 cores BG/L
Wednesday, February 10, 16
Why use message passing
• Allows control over data layout, locality and
communication — very important on large machines
• Portable across all machines including shared memory
machines — it’s a universal parallel programming model
• Easier to write deterministic programs
• simplifies debugging
• easier to understand programs
• Style needed for efficient messages can lead to better
performance than shared memory programs, even on
shared memory systems.
Wednesday, February 10, 16
Why not use it?
• All or nothing program development – generally
need to make the entire program parallel to make
any part parallel
• Information needed for messages low-level and
hard to program
• Subtle bugs in message passing code can lead to
performance problems and deadlock
• Message passing code disrupts the flow of
algorithms
Wednesday, February 10, 16
SPMD execution
• Single Program Multiple Data
• Multiple copies of the same program
operating on different parts of the data
(typically different sections of an array)
• Each program copy executes in a process
• Different processes can execute different
paths through the program
Wednesday, February 10, 16
SPMD execution
for (i=1; i <= n/2; i++) { a[i] = i; } for (i=1, i<= n/2; i++) { ... = a[i-1]; } for (i=1; i <= n/2; i++) { a[i] = n/2+i; } for (i=1; i <= n/2; i++) { ... = a[i-1]; }i 1 0 1 ... n/2-1 n/2 1 2 ... 49 50 a i 1 0 1 ... n/2-1 n/2 51 52 ... 99 100 a n 99n 99 for (i=1; i <= n; i++) { a[i] = i + 1; } for (i=1, i <= n; i++) { ... = a[i-1]; } local index global index Note fixed loop bounds, subscripts and entries in "a" in figure below. Wednesday, February 10, 16 Work done by processes • Each process has a unique rank or process id (often called pid in programs) that is set when program starts • Is not changed during the execution of the program (however, see Naik, Moreira, et al. IBM DRMS project if you are really interested in this.) • Each process has a unique identifier (often called pid) that is known to the program • Typical program pattern is compute ! communicate !compute ... !... !communicate Wednesday, February 10, 16 Radix sort • Radix sort works well to sort lists of numbers • Will assume integers have values from 0 to 65,535 • Have N >> 65,535 numbers to sort
Wednesday, February 10, 16
Sequential program
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}
for (i=0; i < n; i++) {
sorted[data[i]]++;
}
for (i=0; i<65535; i++) {
for (j=0; j < sort[i]; j++) {
fprint(“%i\n”, i);
}}
Want to convert to
SPMD message
passing code
Note that data input not
shown -- this can require
some thought
Data often spread across
multiple files to accommodate
parallel I/O on large problems
Wednesday, February 10, 16
SPMDizing the program
all processors execute this (replicated execution)
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}
each processor executes N/4 iterations (assume N mod 4 = 0)
for (i=0; i < N/4; i++) {
sorted[data[i]]++;
}
this becomes a sum reduction over the sorted arrays on each processor,
i.e. communication. This code does not show that.
for (i=0; i<65535; i++) {
for (j=0; j < sort[i]; j++) {
fprint(“%i\n”, i);
}}
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P3
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P2
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
Data management
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P3
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P2
data[3*N/4:N-1]
i, j
sorted[0:65353]
• All declared variables exist within each process
• There is a global and local logical index space for arrays
• globally, data has N elements 0:N-1
• locally, each process has N/4 elements numbered 0:N/
4-1(if N mod 4 == 0, otherwise !N/4"or#N/4$elements
per processors with some processors having more or
fewer elements than other processors
• The concatenation of the local partitions of data arrays
forms the global array data
• The array data is block distributed over the processors
global indices
shown, local is
[0:n/4-1]
Wednesday, February 10, 16
Data bounds for block
• Two “obvious” ways to compute
• Let n be the array size, P the number
processors
Wednesday, February 10, 16
First method
• Let P be the number of processes, n the number of array elements, 0 ≤ p ≤ P-1 is a
process id
• r = n mod P, r = 0, all blocks are the same size, otherwise, first r blocks have !n/P"
elements, last n-r have #n/P$ elements
• First element on a process p is p⎣n/P⎦+ min(p,r)
• Last element on process p is (p+1)⎣n/P⎦+ min(p+1,r) - 1
• process with element i is min(#i/(#n/P$+ 1)$, #i-r) / #n/P$$)
• Example -- 12 elements over 5 processors, 2 = 12 mod 5
• Example -- 12 elements over 7 processors, 5 = 12 mod 7
Wednesday, February 10, 16
Second method
• First element controlled (or owned) by process p is #p n/P$
(first element and first process id p is 0
• Last element controlled by process p is one less that the
first element controlled by process p+1
# (p+1) n/P$ - 1
• Process controlling element i is #(P(i+1)-1)/n$
• Example -- 12 elements over 5 processors, r = 2 = 12 mod 5
• Example -- 17 elements over 5 processors, r = 2 = 17 mod 5
Wednesday, February 10, 16
Global vs local indices
• Each part of an array within a process must be
indexed as a local element of that array using the local
index.
• Logically, each local element is a part of the global
array, and within the problem domain has a global
index
• It is the MPI programmer’s responsibility (that means
you) to maintain that mapping.
0 1 0 1 20 1 20 10 1
7 8 9 10 114 5 62 30 1
local index:
global index:
Wednesday, February 10, 16
Use macros to access
bounds
• Macros or functions can be used to compute these.
• Block lower bound: LB(p, P, n) = (p*n/P)
• Block upper bound: UB(p, P, n) = LB(p+1, P, n)-1
• Block size: LB(p+1, P, n) - LB(p, P, n)
• Block owner: Owner(i, P, N) = (P*(i+1)-1)/n
0 1 0 1 20 1 20 10 1
7 8 9 10 114 5 62 30 1
local index:
global index:
Wednesday, February 10, 16
Comparison of the two
methods
Operations
First
Method
Second
Method
Low index 4 2
High index 6 4
Owner 7 4
Assumes floor is free (as it is with integer division
although integer division itself may be expensive)
Wednesday, February 10, 16
The cyclic distribution
data[0:N:4]
i, j
sorted[0:65353]
P0 P1
data[1:n:4]
i, j
sorted[0:65353]
P3
data[2:N:4]
i, j
sorted[0:65353]
P2
data[3:N:4]
i, j
sorted[0:65353]
• Let A be an array with N elements.
• Let the array be cyclically distributed over P processes
• Process p gets elements p, p+P, p+2*P, p+3*P, ...
• In the above
• process 0 gets elements 0, 4, 8, 12, ... of data
• process 1 gets elements 1, 5, 9, 13, ... of data
• process 2 gets elements 2, 6, 10, 14, ... of data
• process 3 gets elements 3, 7, 11, 15, ... of data
Wednesday, February 10, 16
The block-cyclic distribution
• Let A be an array with N elements
• Let the array be block-cyclically distributed over P
processes, with blocksize B
• Block b, b = 0 ..., on process p gets elements
b*B*P+p*B: b*B*P + (p+1)*B)-1 elements
• With P=4, B=3
• process 0 gets elements [0:2], [12:14], [24:26] of data
• process 1 gets elements [3:5], [15:17],[27:29] of data
• process 2 gets elements [6:8], [18:20],[30:32] of data
• process 3 gets elements [9:11], [21:23],[33:35] of
data
Wednesday, February 10, 16
Converting the program to MPI: System initialization
#include
#include
// all processors execute this (replicated execution)
int main(int argc, char * argv[ ]) {
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv); // get N from the arg vector
int sorted[65536]; int data[N/4];
MPI_INIT(&argc, &argv);
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}}
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P2
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P3
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
MPI_INIT
• Initialize the MPI runtime
• Does not have to be the first executable statement
in the program, but it must be the first MPI call made
• Initializes the default MPI communicator
(MPI_COMM_WORLD which includes all
processes)
• Reads standard files and environment variables to
get information about the system the program will
execute on
• e.g. what machines executes the program?
Wednesday, February 10, 16
The MPI environment
A
communicator
defines a
universe of
processes that
can exchange
messagesMPI_COMM_WORLD
0
6
1
2
4
3
7
5
A process
A rank
The communicator name
(MPI_COMM_WO LD is the default
communicator name
Wednesday, February 10, 16
Converting the program to MPI
#include
#include
/ all processors execute this (replicated execution)
int main(int argc, char * argv[ ]) {
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv);
int sorted[65536]; int data[N/4];
MPI_INIT(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numP);
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}}
Communicator
name
get
number of
processors
cheat!
should
malloc
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[0:N/4-1]
i, j
sorted[0:65353]
P2
data[0:N/4-1]
i, j
sorted[0:65353]
P3
data[0:N/4-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
Converting the program to MPI
#include
#include
/ all processors execute this (replicated execution)
int main(int argc, char * argv[ ]) {
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv);
int sorted[65536]; int data[*N/4]; MPI_INIT(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numP);
MPI_Comm_rank(MPI_COMM_WORLD, &pid);
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}}
Communicator
name
arg to get
rank (i.e. pid) of
this processor
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P2
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P3
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
Converting the program to MPI
#include
#include
/ all processors execute this (replicated execution)
int main(int argc, char * argv[ ]) {
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv);
int sorted[65536]; int data[*N/4]; MPI_INIT(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numP);
MPI_Comm_rank(MPI_COMM_WORLD, &pid);
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}
MPI_Finalize( );
}
The last MPI
function called
MPI_Finalize frees
system resources
associated with MPI
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P2
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P3
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
Time to do something useful
#include
#include
/ all processors execute this (replicated execution)
int main(int argc, char * argv[ ]) {
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv);
int sorted[65536]; int data[*N/4];
MPI_INIT(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numP);
MPI_Comm_rank(MPI_COMM_WORLD, &pid);
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}
sort(data, sort, pid, numP);
MPI_Finalize( );
}
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P2
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P3
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
The serial code
void sort (sort[ ], data[ ], int pid, int numP) {
for (i=0; i < N; i++) {
sorted[data[i]]++;
}
// sorted results available here ...
}
If above is done in parallel, need to get results from all
processes before printing them
for (i=0; i<65535; i++) {
for (j=0; j < sort[i]; j++) {
fprint(“%i\n”, i);
}}
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P2
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P3
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
MPI_Reduce(...)
• Does a reduction like the reduce clause in
OpenMP, only it uses messages.
MPI_Reduce(void *opnd, void *result, int count, MPI_Datatype type,
MPI_Operator op, int root, MPI_Comm comm);
address
of the first
element to
be reduced
address of the first result
element
number
reduction
elements/
results
type of data
being reduced
reduction
operation
rank of the
process getting
the result
the
communicator over
which the reduction is
performed
Wednesday, February 10, 16
MPI_Datatype
• Defined as constants in the mpi.h header file
• Types supported are
MPI_CHAR MPI_DOUBLE
MPI_FLOAT MPI_INT
MPI_LONG MPI_LONG_DOUBLE
MPI_SHORT MPI_UNSIGNED_CHAR
MPI_UNSIGNED MPI_UNSIGNED_LONG
MPI_UNSIGNED_SHORT
Wednesday, February 10, 16
MPI_Op
• Defined as constants in the mpi.h header file
• Types supported are
MPI_BAND MPI_BOR
MPI_EXOR MPI_BXOR
MPI_LAND MPI_LOR
MPI_LXOR MPI_MAX
MPI_MAXLOC MPI_MIN
MPI_MINLOC MPI_PROD
MPI_SUM
Wednesday, February 10, 16
MPI_Reduce(...)
• Does a reduction like the reduce clause in
OpenMP, only it uses messages.
MPI_Reduce(MPI_IN_PLACE, void *opnd, int count, MPI_Datatype type,
MPI_Operator op, int root, MPI_Comm comm);
address of the first
result element
number
reduction
elements/
results
type of data
being reduced
reduction
operation
rank of
process getting
the result
the
communicator
use
*result as
in and out
buffer on
root
Wednesday, February 10, 16
Example of reduction
MPI_Reduce(MPI_IN_PLACE, sorted, 8, MPI_INT,
MPI_SUM, 0, MPI_COMM_WORLD);
3 5 2 9 8 11 20 4sorted, p=0
8 3 6 8 38 5 27 6sorted, p=1
1 0 9 0 2 1 2 40sorted, p=2
13 15 12 19 18 21 42 3sorted, p=3
25 23 39 36 64 38 91 53sorted, p=0
Wednesday, February 10, 16
Add the reduction
void sort (sort[ ], data[ ], int pid, int numP) {
for (i=0; i < N; i++) {
sorted[data[i]]++;
}
// can merge all of the “sorted” arrays here
if (pid == 0) {
MPI_Reduce(MPI_IN_PLACE, sorted, 8, MPI_INT,
MPI_SUM, 0, MPI_COMM_WORLD);
} else {
MPI_Reduce(sorted, (void *) null, 65353, MPI_INT,
MPI_SUM, 0, MPI_COMM_WORLD);
}
Alternatively, could allocate a buffer for final sorted result. Buffer
would be the same size as sorted.
data[0:N/4-1]
i, j
sorted[0:65353]
P0 P1
data[N/4:2*N/4-1]
i, j
sorted[0:65353]
P2
data[2*N/4:3*N/4-1]
i, j
sorted[0:65353]
P3
data[3*N/4:N-1]
i, j
sorted[0:65353]
Wednesday, February 10, 16
Notes on Reduce
• There is a result for each element of the
source array across all processors
• The result ends up on only one processor
(allreduce sends the result to all
processors)
Wednesday, February 10, 16
Determining program
performance
• MPI_Barrier - barrier
synchronization
• MPI_Wtick - returns
the clock resolution in
seconds
• MPI_Wtime - current
time
int main(int argc, char * argv[ ]) {
...
double elapsed;
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv);
for (i=0; i < 65535; i++) {
sorted[i] = 0;
}
MPI_Barrier( );
elapsed = -MPI_Wtime( );
sort(data, sort, pid, numP);
elapsed += MPI_Wtime( );
if (pid == 0) printSort(final);
MPI_Finalize( );
}
Wednesday, February 10, 16
Determining program
performance
int main(int argc, char * argv[ ]) {
...
double elapsed;
int pid; /* MPI process ID)
int numP; /* number of MPI processes */
int N;
extractArgv(&N, argv); for (i=0; i < 65535; i++) {
sorted[i] = 0;
}
MPI_Barrier( );
elapsed = -MPI_Wtime( );
sort(data, sort, pid, numP);
elapsed += MPI_Wtime( );
if (pid == 0) printSort(final, elapsed);
MPI_Finalize( );
}
Holds the
elapsed time
wait for all
processors to
finish
initialization
negative
of start
time
plus finish
time gives
elapsed time
Wtick( ) returns a
double that holds
the number of
seconds between
clock ticks - 10-3 is
milliseconds
Wednesday, February 10, 16
Wtick( ) gives the clock
resolution
MPI_WTick returns the resolution of MPI_WTime
in seconds. That is, it returns, as a double precision
value, the number of seconds between successive
clock ticks.
double tick = MPI_WTick( );
Thus, a millisecond resolution timer will return 10-3
Wednesday, February 10, 16
Sieve of Erosthenes
• Look at block allocations
• Performance tuning
• MPI_Bcast function
Wednesday, February 10, 16
Finding prime numbers
10987654321
20191817161514131211
30292827262524232221
40393837363534333231
50494847464544434241
60595857565554535251
70696867666564636261
80797877767574737271
90898887868584838281
100999897969594939291
To find primes
1.start with two, mark
all multiples
2.find the next
unmarked u -- it is a
prime
3.mark all multiples of u
between k2 and n until
k2 > n
4.repeat 2 & 3 until
finished
Wednesday, February 10, 16
Finding prime numbers
10987654321
20191817161514131211
30292827262524232221
40393837363534333231
50494847464544434241
60595857565554535251
70696867666564636261
80797877767574737271
90898887868584838281
100999897969594939291
To find primes
3 is prime
mark all multiples of 3 >
9
Wednesday, February 10, 16
Finding prime numbers
10987654321
20191817161514131211
30292827262524232221
40393837363534333231
50494847464544434241
60595857565554535251
70696867666564636261
80797877767574737271
90898887868584838281
100999897969594939291
To find primes
5 is prime
mark all multiples of 5 >
25
Wednesday, February 10, 16
Finding prime numbers
10987654321
20191817161514131211
30292827262524232221
40393837363534333231
50494847464544434241
60595857565554535251
70696867666564636261
80797877767574737271
90898887868584838281
100999897969594939291
To find primes
7 is prime
mark all multiples of 7 >
49
Wednesday, February 10, 16
Finding prime numbers
10987654321
20191817161514131211
30292827262524232221
40393837363534333231
50494847464544434241
60595857565554535251
70696867666564636261
80797877767574737271
90898887868584838281
100999897969594939291
To find primes
11 is prime
mark all multiples of 11
> 121
Wednesday, February 10, 16
Finding prime numbers
10987654321
20191817161514131211
30292827262524232221
40393837363534333231
50494847464544434241
60595857565554535251
70696867666564636261
80797877767574737271
90898887868584838281
100999897969594939291
To find primes
1, 2, 3, 5, 7, 13, 17, 19, 23,
29, 31, 37, 41, 43, 47, 53,
59, 61, 67, 71, 73, 79, 83,
89 and 97 are prime.
1 is not prime by
definition
Wednesday, February 10, 16
Want to parallelize this
• Because we are message passing, obvious
thing to look at it domain decomposition,
i.e. how can we break up the domain being
operated on over multiple processors
• partition data across processors
• associate tasks with data
• In general, try to find fundamental
operations and associate them with data
Wednesday, February 10, 16
What is (are) the
fundamental operation(s)?
• Marking of the
multiples of the
last prime found
• if v a multiple of k
then v mod k == 0
forall (v = k; v < n+1; v++) { if (v mod k != 0) a[v] = 1; } • min-reduction to find the next prime (i.e. smallest unmarked value) across all processes • broadcast the value to all tasks Wednesday, February 10, 16 To make this efficient • Combine as many tasks as possible onto a single process • Make the amount of work done by each process similar, i.e. load balance • Make the communication between tasks efficient Wednesday, February 10, 16 Combining work/ partitioning data • Because processes work on data that they own (the owners compute rule, Rogers and Pingali), the two problems are tightly inter-related. • Each element is owned by a process • It is the process that owns the consistent, i.e., up-to- date value of a variable • All updates to the variable are made by the owner • All requests for the value of the variable are to the owner Wednesday, February 10, 16 Combining work/ partitioning data • Because processes update the data that they own • Cyclic distributions have the property that for all elements i on some process p, i mod p = c where c is some integer value • Although cyclic usually gives better load balance, it doesn’t in this case • Lesson -- don’t apply rules-of-thumb blindly • Block, in this case, gives a better load balance • computation of indices will be harder Wednesday, February 10, 16 Interplay of decomposition and implementation • Decomposition affects how we design the implementation • More abstract issues of parallelization can affect the implementation • In the current algorithm, let Φ be the highest possible prime • At most, only first √Φ values may be used to mark off (sieve) other primes • if P processes, n elements to a process, then if n/P > √ Φ
only elements in p=0 will be used to sieve. This means we only
need to look for lowest unmarked elements in p=0 and only p=0
needs to send this out, saving a reduction operation.
Wednesday, February 10, 16
Use of block partitioning
affects marking
• Can mark j, j+k, j+2k, … where j is the first
prime in the block
• Using the parallel method described in
earlier psuedo-code, would need to use an
expensive mod
for all e in the block
if e mod k = 0, mark e
• We would like to eliminate this.
Wednesday, February 10, 16
Sketch of the algorithm
1. Create list of possible primes
2. On each process, set k = 2
3. Repeat
3.1.On each process, mark all multiples of k
3.2.On process 0, find smallest unmarked number u, set k=u
3.3.On process 0, broadcast k to all processes
4. Until k2 > Φ (the highest possible prime)
5. Perform a sum reduction to determine the number of primes
Wednesday, February 10, 16
Data layout, primes up to 28
2 3 4 5 6 7 8 9 10P=0
0 1 2 3 4 5 6 7 8i =
11 12 13 14 15 16 17 18 19P=1
0 1 2 3 4 5 6 7 8i =
20 21 22 23 24 25 26 2 28P=2
0 1 2 3 4 5 6 7 8i =
array
element
number
being
checked for
“primeness”
Wednesday, February 10, 16
Algorithm 1/4
#include
#include
#include
#include “MyMPI.h”
#define MIN(a,b) ((a)<(b)?(a):(b))
int main (int argc, char *argv[])
{
...
MPI_Init (&argc, &argv);
MPI_Barrier(MPI_COMM_WORLD);
elapsed_time = -MPI_Wtime();
MPI_Comm_rank (MPI_COMM_WORLD, &id);
MPI_Comm_size (MPI_COMM_WORLD, &p);
if (argc != 2) {
if (!id) printf ("Command line: %s
MPI_Finalize(); exit (1);
}
standard
stuff
bounds
macros, etc.
setup,
check args,
etc.
Wednesday, February 10, 16
Algorithm, 2/4
n = atoi(argv[1]);
low_value = 2 + BLOCK_LOW(id,p,n-1);
high_value = 2 + BLOCK_HIGH(id,p,n-1);
size = BLOCK_SIZE(id,p,n-1);
proc0_size = (n-1)/p;
if ((2 + proc0_size) < (int) sqrt((double) n)) {
if (!id) printf ("Too many processes\n");
MPI_Finalize();
exit (1);
}
marked = (char *) malloc (size);
if (marked == NULL) {
printf ("Cannot allocate enough memory\n");
MPI_Finalize();
exit (1);
}
Get min and
max possible
prime on p in
global space
Figure out if
too many
processes for
√Φ candidates
on p=0
allocate array
to use to mark
primes
Wednesday, February 10, 16
BLOCK_LOW
values for P=0, similar for other
processes
11 12 13 14 15 16 17 18 19P=0
9 10 11 12 13 14 15 16 17i =
2 3 4 5 6 7 8 9 10P=0
0 1 2 3 4 5 6 7 8i =
20 21 22 23 24 25 26 2 28P=0
18 19 20 21 22 23 24 25 26i =
low_value
BLOCK_HIGH
high_value
i's are in
global
index
space
Wednesday, February 10, 16
Algorithm 3/4
for (i = 0; i < size; i++) marked[i] = 0; // initialize marking array
if (!id) index = 0; // p=0 action, find first prime
prime = 2;
do { // prime = 2 first time through, sent by bcast on later iterations
if (prime * prime > low_value) // find first value to mark
first = prime * prime – low_value; // first item in this block
else {
if (!(low_value % prime)) first = 0; // first element divisible by prime
else first = prime – (low_value % prime);
}
for (i = first; i < size; i += prime) marked[i] = 1; // mark every kth item
if (!id) { // p=0 action, find next prime by finding unmarked element
while (marked[++index]);
prime = index + 2;
}
MPI_Bcast (&prime, 1, MPI_INT, 0, MPI_COMM_WORLD);
} while (prime * prime <= n);
Wednesday, February 10, 16
First prime index = 0prime = 2
2 3 4 5 6 7 8 9 10P=0
0 1 2 3 4 5 6 7 8local i =
11 12 13 14 15 16 17 18 19P=0
0 1 2 3 4 5 6 7 8local i =
20 21 22 23 24 25 26 2 28P=0
0 1 2 3 4 5 6 7 8local =
2 * 2 > 2
first = 2 * 2 – 2
first = 2
not 2 * 2 > 11
11 % 2 == 1
first = 2 – (l1 % 2)
first = 1
not 2 * 2 > 20
20 % 2 == 0
first = 0
Wednesday, February 10, 16
third prime index = 3prime = 5
2 3 4 5 6 7 8 9 10P=0
0 1 2 3 4 5 6 7 8local i =
11 12 13 14 15 16 17 18 19P=0
0 1 2 3 4 5 6 7 8local i =
20 21 22 23 24 25 26 2 28P=0
0 1 2 3 4 5 6 7 8local =
5 * 5 > 2
first = 5 * 5 – 2
first = 23
5 * 5 > 11
first = 5 * 5 – 11
first = 16
5 * 5 > 20
first = 5 * 5 – 20
first = 5
Wednesday, February 10, 16
Mark every prime elements
starting with first index = 0prime = 2
2 * 2 > 4
first = 2 * 2 – 2
first = 2
not 2 * 2 > 11
11 % 2 == 1
first = 2 – (l1 % 2)
first = 1
not 2 * 2 > 20
20 % 2 == 0
first = 0
2 3 4 5 6 7 8 9 10P=0
0 1 2 3 4 5 6 7 8local i =
11 12 13 14 15 16 17 18 19P=0
0 1 2 3 4 5 6 7 8local i =
20 21 22 23 24 25 26 2 28P=0
0 1 2 3 4 5 6 7 8local =
Wednesday, February 10, 16
Algorithm 4/4
// on each processor count the number of primes, then reduce this total
count = 0;
for (i = 0; i < size; i++)
if (!marked[i]) count++;
MPI_Reduce (&count, &global_count, 1, MPI_INT, MPI_SUM,
0, MPI_COMM_WORLD);
elapsed_time += MPI_Wtime();
if (!id) {
printf ("%d primes are less than or equal to %d\n",
global_count, n);
printf ("Total elapsed time: %10.6f\n", elapsed_time);
}
MPI_Finalize ();
return 0;
}
Wednesday, February 10, 16
Mark every prime elements
starting with first index = 0prime = 2
2 3 4 5 6 7 8 9 10P=0
11 12 13 14 15 16 17 18 19P=0
20 21 22 23 24 25 26 27 28P=0
count = 1
count = 4
count = 2
global_count = 1 + 4 + 2
Wednesday, February 10, 16
Other MPI environment
management routines
• MPI_Abort (comm, errorcode)
• Aborts all processors associated with communicator
comm
• MPI_Get_processor_name(&name, &length)
• MPI version of gethostname, but what it returns is
implementation dependent. gethostname may be
more portable.
• MPI_Initialized(&flag)
• Returns true if MPI_Init has been called, false
otherwise
Wednesday, February 10, 16
point-to-point communication
• Most MPI communication is between a pair of
processors
• send/receive transmits data from the sending process
to the receiving process
• MPI point-to-point communication has many flavors:
• Synchronous send
• Blocking send / blocking receive
• Non-blocking send / non-blocking receive
• Buffered send
• Combined send/receive
• "Ready" send (matching receive already posted.)
• All types of sends can be paired with all types of receive
Wednesday, February 10, 16
Buffering
What happens when
• A send occurs before the receiving process is
ready for the data
• The data from multiple sends arrive at the
receiving task which can only accept one at a time
Wednesday, February 10, 16
System buffer space
Not part of the standard -- an “implementation detail
• Managed and controlled by the MPI library
• Finite
• Not well documented -- size maybe a function of
install parameters, consequences of running out not
well defined
• Both sends and receives can be buffered
• Can help performance by allowing asynchronous
send/recvs
• Can hurt performance because of memory copies
• Program variables are called application buffers in MPI-
speak
Wednesday, February 10, 16
Blocking and non-blocking point-to-
point communication
Blocking
• Most point-to-point routines have a blocking and non-blocking mode
• A blocking send call returns only when it is safe to modify/reuse the application
buffer. Basically the data in the application buffer has been copied into a system
buffer or sent.
• Blocking send can be synchronous, which means call to send returns when data is
safely delivered to the recv process
• Blocking send can be asynchronous by using a send buffer
• A blocking receive call returns when sent data has arrived and is ready to use
• Non-blocking
• Non-blocking send and receive calls behave similarly and return almost
immediately.
• Non-blocking operations request the MPI library to perform the operation when it
is able. It cannot be predicted when the action will occur.
• You should not modify any application buffer (program variable) used in non-
blocking communication until the operation has finished. Wait calls are available to
test this.
• Non-blocking communication allows overlap of computation with communication
to achieve higher performance
Wednesday, February 10, 16
Synchronous and buffered
sends and receives
• synchronous send operations block until the receiver
begins to receive the data
• buffered send operations allow specification of a buffer
used to hold data (this buffer is not the application buffer
that is the variable being sent or received)
• allows user to get around system imposed buffer limits
• for programs needing large buffers, provides portability
• One buffer/process allowed
• synchronous and buffered can be matched
Wednesday, February 10, 16
Ordering of messages and fairness
• Messages received in-order
• If a sender sends two messages, (m1 and m2) to the same
destination, and both match the same receive, m1 will be
received before m2.
• If a receiver posts two receives (r1 followed by r2), and
both are looking for the same messages, r1 will receive a
message before r2.
• Operation starvation is possible
• task2 performs a single receive. task0 and task3 both
send a message to task2 that matches the receive. Only
one of the sends will complete if the receive is only
executed once.
• It is the programmer’s job to ensure this doesn’t happen
Wednesday, February 10, 16
Operation starvation
Only one of the sends will
complete.
Networks are generally not
deterministic, cannot be
predicted whose message
will arrive at task2 first, and
which will complete.
Wednesday, February 10, 16
Basic sends and receives
• MPI_send(buffer, count, type, dest, tag, comm)
• MPI_Isend(buffer, count, type, dest, tag, comm,
request)
• MIP_Recv(buffer, count, type, source, tag, comm,
status)
• MPI_Irecv(buffer, count, type, source, tag, comm,
request)
I forms are non-blocking
Wednesday, February 10, 16
Basic sends/recv arguments (I forms are
non-blocking)
• MPI_send(buffer, count, type, dest, tag, comm)
• MPI_Isend(buffer, count, type, dest, tag, comm, request)
• MIP_Recv(buffer, count, type, source, tag, comm, status)
• MPI_Irecv(buffer, count, type, source, tag, comm,
request)
• buffer: pointer to the data to be sent or where
received (a program variable)
• count: number of data elements (not bytes!) to be sent
• type: an MPI_Type
• tag: the message type, any unsigned integer 0 - 32767.
• comm: sender and receiver communicator
Wednesday, February 10, 16
Basic send/recv arguments
• MPI_send(buffer, count, type, dest, tag, comm)
• MPI_Isend(buffer, count, type, dest, tag, comm, request)
• MIP_Recv(buffer, count, type, source, tag, comm, status)
• MPI_Irecv(buffer, count, type, source, comm, request)
• dest: rank of the receiving process
• source: rank of the sending process
• request: for non-blocking operations, a handle to an
MPI_Request structure for the operation to allow wait type
commands to know what send/recv they are waiting on
• status: the source and tag of the received message. This is a
pointer to the structure of type MPI_Status with fields
MPI_SOURCE and MPI_TAG.
Wednesday, February 10, 16
Blocking send/recv/etc.
MPI_Send: returns after buf is free to be reused. Can use a system buffer
but not required, and can be implemented by a system send.
MPI_Recv: returns after the requested data is in buf.
MPI_Ssend: blocks sender until the application buffer is free and the
receiver process started receiving the message
MPI_Bsend: permits the programmer to allocate buffer space instead of
relying on system defaults. Otherwise like MPI_Send.
MPI_Buffer_attach (&buffer,size): allocate a message buffer with the
specified size
MPI_Buffer_detach (&buffer,size): frees the specified buffer
MPI_Rsend: blocking ready send, copies directly to the receive application
space buffer, but the receive must be posted before being invoked.
MPI_Sendrecv: performs a blocking send and a blocking receive. Processes
can swap without deadlock
Wednesday, February 10, 16
Wait and probe
MPI_Wait (&request, &status): wait until the operation specified by
request (specified in an Isend/Irecv finishes)
MPI_Waitany (count, &array_of_requests, &index,&status): wait
for any blocking operations specified in &array_of_requests to finish
MPI_Waitall (count, &array_of_requests, &array_of_statuses): wait
for all blocking operations specified in &array_of_requests to finish
MPI_Waitsome (incount, &array_of_requests, &outcount,
&array_of_offsets, &array_of_statuses): wait for at least one request
to finish, the number is returned in outcount.
MPI_Probe (source, tag, comm, &status): performs a blocking test
but doesn’t require a corresponding receive to be posted.
Wednesday, February 10, 16
Example of blocking send/recv
#include "mpi.h"
#include
int main(argc,argv)
int argc;
char *argv[]; {
int numtasks, rank, dest, source, rc, count, tag=1;
char inmsg, outmsg=’x’;
MPI_Status Stat; // status structure
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
Wednesday, February 10, 16
Example of blocking send/recv
if (rank == 0) {
dest = 1;
source = 1;
rc = MPI_Send(&outmsg, 1, MPI_CHAR, dest, tag, MPI_COMM_WORLD);
rc = MPI_Recv(&inmsg, 1, MPI_CHAR, source, tag, MPI_COMM_WORLD, &Stat);
} else if (rank == 1) {
dest = 0;
source = 0;
rc = MPI_Recv(&inmsg, 1, MPI_CHAR, source, tag, MPI_COMM_WORLD, &Stat);
rc = MPI_Send(&outmsg, 1, MPI_CHAR, dest, tag, MPI_COMM_WORLD);
}
rc = MPI_Get_count(&Stat, MPI_CHAR, &count); // returns # of type received
printf(“Task %d: Received %d char(s) from task %d with tag %d \n”,
rank, count, Stat.MPI_SOURCE, Stat.MPI_TAG);
MPI_Finalize( );
}
Wednesday, February 10, 16
Example of blocking send/recv
if (rank == 0) {
dest = 1;
source = 1;
rc = MPI_Send(&outmsg, 1, MPI_CHAR, dest, tag, MPI_COMM_WORLD);
rc = MPI_Recv(&inmsg, 1, MPI_CHAR, source, tag, MPI_COMM_WORLD, &Stat);
} else if (rank == 1) {
dest = 0;
source = 0;
rc = MPI_Recv(&inmsg, 1, MPI_CHAR, source, tag, MPI_COMM_WORLD, &Stat);
rc = MPI_Send(&outmsg, 1, MPI_CHAR, dest, tag, MPI_COMM_WORLD);
}
task0 task1
green/italic send
blue/bold send
Wednesday, February 10, 16
Why the reversed send/recv orders?
if (rank == 0) {
dest = 1;
source = 1;
rc = MPI_Send(&outmsg, 1, MPI_CHAR, dest, tag, MPI_COMM_WORLD);
rc = MPI_Recv(&inmsg, 1, MPI_CHAR, source, tag, MPI_COMM_WORLD, &Stat);
} else if (rank == 1) {
dest = 0;
source = 0;
rc = MPI_Recv(&inmsg, 1, MPI_CHAR, source, tag, MPI_COMM_WORLD, &Stat);
rc = MPI_Send(&outmsg, 1, MPI_CHAR, dest, tag, MPI_COMM_WORLD);
}
MPI_Send may or may not block. It will block until the sender
can reuse the sender buffer. Some implementations will return to
the caller when the buffer has been sent to a lower communication
layer. Some others will return to the caller when there’s a matching
MPI_Recv() at the other end. So it’s up to your MPI
implementation whether if this program will deadlock or not.
From stackoverflow http://stackoverflow.com/questions/20448283/deadlock-with-mpi
Wednesday, February 10, 16
http://stackoverflow.com/questions/20448283/deadlock-with-mpi
http://stackoverflow.com/questions/20448283/deadlock-with-mpi
Non-blocking operations
• MPI_Isend, MPI_Irecv, MPI_Issend, Ibsend, Irsend: similar
to MPI_Send, MPI_Recv, MPI_Ssend, Bsend, Rsend except
that a Test or Wait must be used to determine that the
operation has completed and the buffer may be read (in the
case of a recv) or written (in the case of a send).
• MPI_Test (&request, &flag,&status)
• MPI_Testany (count, &array_of_requests, &index, &flag, &status)
• MPI_Testall (count,&array_of_requests,&flag, &array_of_statuses)
• MPI_Testsome (incount, &array_of_requests, &outcount,
&array_of_offsets, &array_of_statuses)
• Like the wait operations, but do not block
Wednesday, February 10, 16
Non-blocking example
#include “mpi.h”
#include
int main(argc,argv)
int argc;
char *argv[]; {
int numtasks, rank, next, prev, buf[2], tag1=1, tag2=2;
MPI_Request reqs[4];
MPI_Status stats[4];
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
Wednesday, February 10, 16
Non-blocking
example
prev = rank-1;
next = rank+1;
if (rank == 0) prev = numtasks – 1;
if (rank == (numtasks – 1)) next = 0;
MPI_Irecv(&buf[0], 1, MPI_INT, prev, tag1, MPI_COMM_WORLD, &reqs[0]);
MPI_Irecv(&buf[1], 1, MPI_INT, next, tag2, MPI_COMM_WORLD, &reqs[1]);
MPI_Isend(&rank, 1, MPI_INT, prev, tag2, MPI_COMM_WORLD, &reqs[2]);
MPI_Isend(&rank, 1, MPI_INT, next, tag1, MPI_COMM_WORLD, &reqs[3]);
{ do some work }
MPI_Waitall(4, reqs, stats);
MPI_Finalize();
}
Nearest neighbor exchange
in a ring topology
Wednesday, February 10, 16
Collective communication
routines
• Use these when communicating among processes with a well
defined pattern
• Some can be used to allow all processes to communicate
• Some perform computation during the communication
(reductions)
• Involve all processes in the specified communicator, even if a
particular processor has no data to send
• Can only be used with MPI predefined types, not derived
types.
• The programmer has to make sure all processes participate
in the collective operation
Wednesday, February 10, 16
All processors participate
in the collective operation
if (pid % 2) {
MPI_Reduce(…, MPI_COMM_WORLD);
}
This program will deadlock, as the MPI_Reduce
will wait forever for even processes to begin
executing it.
If you want to only involve odd processes, add
them to a new communicator.
Wednesday, February 10, 16
Groups and communicators
• Two terms used in MPI documentation are
groups and communicators.
• A communicator is a group of processes that
can communicate with each other
• A group is an ordered set of processes
• Programmers can view groups and
communicators as being identical
Wednesday, February 10, 16
Collective routines
MPI_Barrier (comm): tasks block upon reaching the barrier until every task in the
group has reached it
MPI_Bcast (&buffer,count,datatype,root,comm): process root sends a copy of its
data to every other processor. Should be log2(comm_size) operation.
MPI_Scatter (&sendbuf,sendcnt,sendtype,&recvbuf,
recvcnt,recvtype,root,comm): distributes a unique message from root to every
process in the group.
MPI_Gather(&sendbuf, sendcnt, sendtype, &recvbuf, recvcount, recvtype,
root, comm): opposite of scatter, every process in the group sends a
unique message to the root.
MPI_Allgather (&sendbuf,sendcount,sendtype,&recvbuf,
recvcount,recvtype,comm): each tasks performs a one-to-all broadcast to every
other process in the group These are concatenated together in the recvbuf.
MPI_Reduce (&sendbuf,&recvbuf,count,datatype,op,root,comm): performs a
reduction using operation op and places the result into recvbuf on the root process.
Wednesday, February 10, 16
MPI_Bcast
Wednesday, February 10, 16
MPI_Scatter
MPI_Send(sendbuf+i*sendcount*extent(sendtype), sendcount, sendtype, i, …)
MPI_Recv(recvbuf, recvcount, recvtype, i, sendcount, sendtype, i, …)
Equivalent to
Wednesday, February 10, 16
MPI_Gather
MPI_Send(sendbuf, sendcount,
sendtype, root, …)
MPI_Recv(recvbuf+
i*recvcount*
extent(recvtype),
recvcount,
recvtype, i, …)
With the results of each
recv stored in rank order of
the sending process
Wednesday, February 10, 16
MPI_Allgather
An gather with
every process
being a target.
Wednesday, February 10, 16
MPI_Reduce
Also see MPI
introductory
slides.
You can form
your own
reduction
function using
MPI_Op_create
Wednesday, February 10, 16
MPI_Op_create
#include “mpi.h”
int MPI_Op_create(MPI_User_function *function, int commute, MPI_Op *op )
pointer
to the user
defined
function
true if
commutative, false
otherwise
Handle
to refer to the
function wherever
an MPI_Op is
needed
Wednesday, February 10, 16
More operations
MPI_Allreduce (&sendbuf, &recvbuf, count, datatype, op, comm): functionally
equivalent to an MPI_Reduce followed by an MPI_Bcast. Faster on most hardware than the
combination of these.
MPI_Reduce_scatter(&sendbuf, &recvbuf, recvcount, datatype, op, comm): Does an
element-wise reduce on the vector in sendbuf of length recvcount. The vector is then split
into disjoint segments and spread across the tasks. Equivalent to an MPI_Reduce followed
by an MPI_Scatter operation.
MPI_Alltoall(&sendbuf, sendcount, sendtype, &recvbuf, recvcnt, recvtype, comm):
Each task in the group performs a scatter with the results concatenated on each process in
task rank order.
MPI_Scan(&sendbuf, &recvbuf, count, datatype, op, comm): performs the partial sums
on each processor that would result from doing an in-order reduction across the
processors in rank order.
Wednesday, February 10, 16
MPI_Allreduce
Wednesday, February 10, 16
P0 P1 P2 P3 P4 P5 P6 P7
P0 P2 P4 P8
P0 P4
P0
P0 P4
P0 P2 P4 P6
P0 P1 P2 P3 P4 P5 P6 P7
0:1 2:3 4:5 6:7
0:3 4:7
0:7
0:7 0:7
0:7 0:7 0:7 0:7
all have 0:7
Naive Allreduce
2*log2(|P|)
steps
Wednesday, February 10, 16
P0 P1 P2 P3 P4 P5 P6 P7
P0 P2 P4 P8
P0 P4
P0
P0 P4
P0 P2 P4 P6
P0 P1 P2 P3 P4 P5 P6 P7
0:1 2:3 4:5 6:7
0:3 4:7
0:7
0:7 0:7
0:7 0:7 0:7 0:7
all have 0:7
Why is this naive? On average
only ~1/2 of nodes involved in
communication each step
8
4
2
2
4
8
Wednesday, February 10, 16
P0 P1 P2 P3 P4 P5 P6 P7
P0 P1 P2 P3 P4 P5 P6 P7
P0 P1 P2 P3 P4 P5 P6 P7
P0 P1 P2 P3 P4 P5 P6 P7
0:1 2:30:1 2:3 4:5 4:5 6:7 6:7
0:3 0:3 0:3 0:3 4:7 4:7 4:7 4:7
0:7 0:7 0:7 0:7 0:7 0:7 0:7 0:7
log2(|P|) steps
Wednesday, February 10, 16
P0 P1 P2 P3 P4 P5 P6 P7
P0 P1 P2 P3 P4 P5 P6 P7
P0 P1 P2 P3 P4 P5 P6 P7
P0 P1 P2 P3 P4 P5 P6 P7
The faster algorithm relies on current network interface cards
being at least dual ported. Each node in the system can
simultaneously send and receive a message.
Algorithm from Optimization of Collective Reduction
Operations, Rolf Rabenseifner, International Conference on
Computational Science, 2004
All processors all
busy each step.
Note that the
bandwidth
requirements of
the network
change
Wednesday, February 10, 16
MPI_Reduce_scatter
0
4
8
12
reduce
result
result of scattering
the reduce result
Wednesday, February 10, 16
MPI_Alltoall
Each process performs
a scatter of its
elements to all other
processes.
Received data is
concatenated in
sender rank order
Wednesday, February 10, 16
MPI_Scan
0 0:1 0:2 0:3
Wednesday, February 10, 16
Group and communicator
• Remember that
• A communicator is a group of processes
that can communicate with each other
• A group is an ordered set of processes
• Programmers can view groups and
communicators as being the same thing
• group routines are used in collecting
processes to form communicator.
Wednesday, February 10, 16
Why groups and communicators?
• Allow programmer to organize tasks by
functions
• Enable collective communication operations
• Allow user-defined virtual topologies to be
formed
• Enable manageable communication by
enabling synchronization
Wednesday, February 10, 16
Properties
• Groups/communicators are dynamic, i.e.
they can be created and destroyed
• Processes can be in many groups, and will
have a unique, possibly different, rank in
each group
• MPI provides 40+ routines for managing
groups and communicators! Mercifully, we
will not cover them all.
Wednesday, February 10, 16
Tasks these 40+ routines can
perform
Extract handle of a global group a communicator using
MPI_Comm_group
• Form new group as a subset of another group using
MPI_Group_incl
• Create new communicator for a group using
MPI_Comm_create
• Determine a processor’s rank in a communicator using
MPI_Comm_rank
• Communicate among the processors of a group
• When finished, free communicators and groups using
MPI_Comm_free and MPI_Group_free
Wednesday, February 10, 16
Relationships among
communicators and
groups.
Both collective
and point-to-point
communication is
within a group.
Wednesday, February 10, 16
#include “mpi.h”
#include
#define NPROCS 8
int main(argc,argv)
int argc;
char *argv[]; {
int rank, new_rank, sendbuf, recvbuf, numtasks,
ranks1[4]={0,1,2,3}, ranks2[4]={4,5,6,7};
MPI_Group orig_group, new_group;
MPI_Comm new_comm;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
if (numtasks != NPROCS) {
printf(“Must specify MP_PROCS= %d. Terminating.\n”,NPROCS);
MPI_Finalize();
exit(0);
}
Handle for
MPI_COMM_WORLD
Handle for a
new group
Handle for a new
communicator
Get the
number of tasks and
the rank of
MPI_COMM_WORLD
sanity check code
Wednesday, February 10, 16
#include “mpi.h”
#include
#define NPROCS 8
int main(argc,argv)
int argc;
char *argv[]; {
int rank, new_rank, sendbuf, recvbuf, numtasks,
ranks1[4]={0,1,2,3}, ranks2[4]={4,5,6,7};
MPI_Group orig_group, new_group;
MPI_Comm new_comm;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
if (numtasks != NPROCS) {
printf(“Must specify MP_PROCS= %d. Terminating.\n”,NPROCS);
MPI_Finalize();
exit(0);
}
Variables to hold information about
the new group this will be in. Note that
since this is an SPMD program, if we do
this statically we need information for
all groups the process can be in, not just
the one that it is in.
Hold the ranks of processors in
(in MPI_COMM_WORLD) of
processes in each of the two new
groups.
Wednesday, February 10, 16
sendbuf = rank;
/* Extract the original group handle */
MPI_Comm_group(MPI_COMM_WORLD, &orig_group);
/* Divide tasks into two distinct groups based upon rank */
if (rank < NPROCS/2) {
MPI_Group_incl(orig_group, NPROCS/2, ranks1, &new_group);
}
else {
MPI_Group_incl(orig_group, NPROCS/2, ranks2, &new_group);
}
/* Create new new communicator and then perform collective communications */
MPI_Comm_create(MPI_COMM_WORLD, new_group, &new_comm);
MPI_Allreduce(&sendbuf, &recvbuf, 1, MPI_INT, MPI_SUM, new_comm);
MPI_Group_rank (new_group, &new_rank);
printf("rank= %d newrank= %d recvbuf= %d\n",rank,new_rank,recvbuf);
MPI_Finalize();
}
get handle for
MPI_COMM_WORLD
Each
process executes one
of these statements.
Based on its number, becomes a
member of one of the new
groups.
Wednesday, February 10, 16
sendbuf = rank;
/* Extract the original group handle */
MPI_Comm_group(MPI_COMM_WORLD, &orig_group);
/* Divide tasks into two distinct groups based upon rank */
if (rank < NPROCS/2) {
MPI_Group_incl(orig_group, NPROCS/2, ranks1, &new_group);
}
else {
MPI_Group_incl(orig_group, NPROCS/2, ranks2, &new_group);
}
/* Create new new communicator and then perform collective communications */
MPI_Comm_create(MPI_COMM_WORLD, new_group, &new_comm);
MPI_Allreduce(&sendbuf, &recvbuf, 1, MPI_INT, MPI_SUM, new_comm);
MPI_Group_rank (new_group, &new_rank);
printf("rank= %d newrank= %d recvbuf= %d\n",rank,new_rank,recvbuf);
MPI_Finalize();
}
Create a
communicator from the
group formed above
Perform collective
communication within the
group
Get the
processes rank
within the new
group
Wednesday, February 10, 16