Commit 0f6e0812 authored by atuncer @kelenken's avatar atuncer @kelenken
Browse files

added monte carlo example

parent 346b4116
......@@ -24,6 +24,7 @@ set(DWARF_PREFIX 8_io) # The prefix of the name of the binaries produced
# Add the examples
add_subdirectory(basicMPIIO)
add_subdirectory(read2shmem)
add_subdirectory(markovchain)
# ==================================================================================================
......
# Packages are optional: if they are not present, certain code samples are not compiled
cmake_minimum_required(VERSION 2.8.10 FATAL_ERROR)
find_package(MPI) # Built-in in CMake
include(${CMAKE_CURRENT_SOURCE_DIR}/../../cmake/common.cmake)
# ==================================================================================================
if ("${DWARF_PREFIX}" STREQUAL "")
set(DWARF_PREFIX 8_io)
endif()
set(NAME ${DWARF_PREFIX}_markovchain)
# C compiler settings
find_package(Common)
if (MPI_FOUND)
cmake_policy(SET CMP0003 OLD)
set (CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
include_directories(${MPI_INCLUDE_PATH})
add_executable(${NAME} markovchain)
target_link_libraries(${NAME} ${MPI_LIBRARIES} stdc++)
install(TARGETS ${NAME} DESTINATION bin)
message("** Enabling '${NAME}': with MPI")
else()
message("## Skipping '${NAME}': no MPI support found")
# dummy_install(${NAME} "MPI")
endif()
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${C_FLAGS}")
unset(NAME)
# ==================================================================================================
# README - "Markov Chain": Analysis of a state sequence to extract history dependence
## Description
This example takes a state sequence from a specified file, and extracts patterns of state transitions.
## Contributors
* Ahmet Tuncer Durak (UHeM / Turkey)
## Copyright
This code is available under Apache License, Version 2.0 - see also the license file in the CodeVault root directory.
## Languages
This sample is entirely written in C.
## Parallelisation
This sample uses MPI-2 for parallelisation.
## Level of the code sample complexity
Advanced
## Compiling
Follow the compilation instructions given in the main directory of the kernel samples directory (`/hpc_kernel_samples/README.md`).
## Running
To run the program, use something similar to:
mpirun -n [nprocs] ./8_io_markovchain [inputfile]
Note that `inputfile` is simply treated as stream of characters representing different states.
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include "mpi.h"
// adjust BITS_PER_STATE and STATES_PER_PATTERN according to the problem
// NOTE: (BITS_PER_STATE * STATES_PER_PATTERN) should not exceed 30 bits
// number of bits to represent a state
#define BITS_PER_STATE 3
// number of states to represent a pattern
#define STATES_PER_PATTERN 4
// number of unique states
#define NUM_STATES (1 << BITS_PER_STATE)
// number of unique patterns
#define NUM_PATTERNS (1 << (BITS_PER_STATE * STATES_PER_PATTERN))
// number of preceding states that result in the last state o the pattern
// *history: extra states to read before establishing first pattern
#define LEN_HISTORY (STATES_PER_PATTERN - 1)
// default state for unknown state symbols
#define UNKNOWN_STATE 0
// bitmask where only bits for a single state is set
#define MASK_STATE (NUM_STATES - 1)
// bitmask where only bits for a complete pattern is set
#define MASK_PATTERN (NUM_PATTERNS - 1)
uint32_t stateID(char c) {
// get the state id for this state symbol
// if this is an unknown symbol, return UNKNOWN_STATE (i. e. 0)
if (c < 'a') return UNKNOWN_STATE;
if (c >= ('a' + NUM_STATES)) return UNKNOWN_STATE;
// return 1 for a, 2 for b, ...
return (c - 'a' + 1);
}
char stateSYM(uint32_t s) {
// get the state symbol for this state id
// truncate overflowing bits (if any) from state
s = s & MASK_STATE;
// if this is an unknown symbol, return '_'
if (s == UNKNOWN_STATE) return '_';
// return a for 1, b for 2, ...
return ('a' + s - 1);
}
uint32_t add_state(uint32_t pattern, uint32_t state) {
// drop oldest state in pattern
// shift pattern by one state
// place state as the newest state in pattern
// truncate overflowing bits (if any) from state
state = state & MASK_STATE;
// shift the pattern to clear space for the newest state
pattern = pattern << BITS_PER_STATE;
// truncate the pattern to drop the oldest state
pattern = pattern & MASK_PATTERN;
// add the given state as the newest state in the pattern
pattern = pattern | state;
// return the altered pattern
return pattern;
}
char * print_pattern(char *buf, uint32_t pattern) {
// print the given pattern onto the given buffer
// for each state in pattern
for (int i=0; i<STATES_PER_PATTERN; i++) {
// print state for the current symbol (moving right to left)
buf[STATES_PER_PATTERN-i-1] = stateSYM(pattern & MASK_STATE);
// drop the newest symbol (moving new to old)
pattern = pattern >> BITS_PER_STATE;
}
// shift final state
buf[STATES_PER_PATTERN] = buf[STATES_PER_PATTERN-1];
// put a seperator before final state
buf[STATES_PER_PATTERN-1] = ':';
// terminate the string
buf[STATES_PER_PATTERN+1] = '\0';
// return buffer
return buf;
}
char *buffer_from_file(char *filename, int *bytes_read) {
// creates a buffer on each process
// reads states (single char symbols) from file (named filename) into buffer
// sets bytes_read to the number of states read into buffer
MPI_File filehandle; // handle for referring to opened file
char *buffer; // buffer to hold states read from file
int id_me; // id of this process
int id_max; // number of processes
MPI_Offset size_file; // number of states in the file
MPI_Offset size_chunk; // number of states per process (excluding history*)
MPI_Offset size_buffer; // number of states the buffer can hold
MPI_Offset size_skip; // number of states to skip before reading into buffer
MPI_Status status; // status to hold the result of the reading attempt
// get the id of this process
MPI_Comm_rank(MPI_COMM_WORLD, &id_me);
// get the number of processes
MPI_Comm_size(MPI_COMM_WORLD, &id_max);
if (MPI_File_open( // open the file ...
MPI_COMM_WORLD, // ... for all processes in MPI_COMM_WORLD
filename, // ... named filename
MPI_MODE_RDONLY, // ... as read only
MPI_INFO_NULL, // ... using default hints (see man page)
&filehandle // ... and refer to it as filehandle
) != MPI_SUCCESS) {
// if the operation is not successfull
printf("ERROR: could not read from input file\n");
return NULL;
}
// get the number of states in the file
MPI_File_get_size(filehandle, &size_file);
// split the number of states (excluding states) to processes
size_chunk = (size_file - LEN_HISTORY) / id_max;
// each buffer should be able to hold a chunk AND history*
size_buffer = size_chunk + LEN_HISTORY;
// the last process should be able to hold up to id_max excess states
// excess states come from the remainder of "file / process" division,
// ... thus, may nox exceed id_max
if (id_me == (id_max - 1)) size_buffer += id_max;
// get the number of states to skip before starting reading
size_skip = id_me * size_chunk;
// allocate memory for buffer
buffer = (char *) malloc(sizeof(char) * ((size_t) size_buffer));
if (buffer == NULL) {
// if the allocation fails
printf("ERROR: could not allocate memory for buffer\n");
return NULL;
}
MPI_File_seek( // reposition the cursor in file ...
filehandle, // ... using filehandle
size_skip, // ... skipping size_skip states
MPI_SEEK_SET); // ... using absolute position (in contrast to relative position) in file
MPI_File_read( // read states from file ...
filehandle, // ... using filehandle
buffer, // ... into buffer
size_buffer, // ... up to size_buffer states
MPI_BYTE, // ... where each state is a MPI_BYTE
&status); // ... recording the result in status variable
// get the number of states actually read from file
MPI_Get_count(&status, MPI_BYTE, bytes_read);
// place a sentinel value at the end of the buffer
buffer[*bytes_read] = '\0';
// close the file
MPI_File_close(&filehandle);
// return the address of the buffer
return buffer;
}
unsigned int *hitcounts_from_buffer(char *buffer, int bytes_read) {
// allocate a hit table for patterns
// scan buffer and the count occurence of each pattern
unsigned int *hits; // number of hits (indexed by pattern)
uint32_t pattern; // current pattern
// allocate memory for the hit table
hits = (unsigned int *) malloc(sizeof(unsigned int) * NUM_PATTERNS);
if (hits == NULL) {
// if the allocation fails
printf("ERROR: could not allocate memory for hit table\n");
return NULL;
}
// initialize the hit table
for (int i=0; i<NUM_PATTERNS; i++) hits[i] = 0;
// start with an empty pattern
pattern = 0;
// fill the pattern to obtain history*
for (int i=0; i<LEN_HISTORY; i++) {
// drop oldest state, add next state from buffer
pattern = add_state(pattern, stateID(buffer[i]));
}
// read states from buffer and update hit table
for (int i=LEN_HISTORY; i<bytes_read; i++) {
// drop oldest state, add next state from buffer
pattern = add_state(pattern, stateID(buffer[i]));
// update the number of times this pattern is encountered
hits[pattern] += 1;
}
// return hi table
return hits;
}
void analyse_and_print_results(unsigned int * hits) {
// gather hit counts from all processes
// calculate frequencies
// print results
int id_me; // id of this process
unsigned int hit_count; // hit count per history
uint32_t pattern; // current pattern
char pbuf[32]; // buffer for printing pattern
// get the id of this process
MPI_Comm_rank(MPI_COMM_WORLD, &id_me);
MPI_Reduce( // reduce local results ...
id_me ? hits : MPI_IN_PLACE, // ... from hits (special value MPI_IN_PLACE for root process)
// ... as this process overwrites its send buffer
hits, // ... into hits
NUM_PATTERNS, // ... for NUM_PATTERNS items
MPI_UNSIGNED, // ... of unsigned int data type
MPI_SUM, // ... appling sum operation
0, // ... gathering in the root process
MPI_COMM_WORLD); // ... for all processes in MPI_COMM_WORLD
if (id_me == 0) { // only on the main process
for (int i=0; i < (1 << (BITS_PER_STATE * LEN_HISTORY)); i++) { // for each history pattern
// reset hit count for this history
hit_count = 0;
for (int j=0; j < NUM_STATES; j++) { // for each resulting state
// count the occurrent of this history pattern
hit_count += hits[(i << BITS_PER_STATE) + j];
}
if (hit_count) { // if this history pattern occured at least once
for (int j=0; j < NUM_STATES; j++) { // for each resulting state
pattern = (i << BITS_PER_STATE) + j; // calculate pattern to avoid repetition
if (hits[pattern]) { // only if hits for this pattern is non-zero
printf("%s : %d / %d\n", print_pattern(pbuf, pattern), hits[pattern], hit_count);
}
}
// insert an empty line to separate history patterns
printf("\n");
}
}
}
}
int main(int argc, char *argv[]) {
unsigned int *hits; // hit counts per pattern
char *buffer; // buffer to read from file
int bytes_read; // number of states read
// initiaize MPI subsystem
MPI_Init(&argc, &argv);
// print an error if the input file is not specified
if (argc != 2) {
printf("USAGE: %s <input_file>\n", argv[0]);
MPI_Finalize();
exit(EXIT_FAILURE);
}
// read a part of the input file into local buffer
buffer = buffer_from_file(argv[1], &bytes_read);
// count occurences of patterns i local buffer
hits = hitcounts_from_buffer(buffer, bytes_read);
// free buffer as it is no longer needed
free(buffer);
// combine and process the data, then print the results
analyse_and_print_results(hits);
// free hits as it is no longer needed
free(hits);
// finalize the MPI subsystem
MPI_Finalize();
// finalize the program
return EXIT_SUCCESS;
}
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment