triangle_reflection_complex/parallelization/parallel.h

119 lines
4.0 KiB
C

#ifndef PARALLEL_H
#define PARALLEL_H
/*
this is a library to parallelize workloads which can be split up naturally
into a sequence of independent jobs, using MPI. A program will usually
- do precomputation
- fill array with input data
- do the parallel work
- print the output data
we want to enable restarts, so that only unfinshed jobs need to be repeated.
Further, we want to be resilient to slow/unreliable network and to losing
nodes. There is a main node and a number of workers. The main node does the
precomputation and then retires do do administrative work, and the workers
do the actual jobs. We also want to switch to serial mode if the program is
called without MPI.
The following data has to be transimitted between nodes:
- results of the precomputation (read-only, shared between nodes)
- job-specific input data, generated by main node before parallel part
- output data for each job
the parallel work shall be given as a callback function which takes
input data and precomputation data as parameter
the above program will look like this for us:
- parallel_init
- if we are a worker, do parallel_work(init_callback, job_callback), exit
- do precomputation
- fill array with input data
- output_array = parallel_run(input_array)
- print the output data
parallel_init:
- check if we're running as an mpi program
- init mpi, check what kind of node we are
parallel_work(init_callback1, init_callback2, job_callback):
- receive global_precomp (???)
- worker_precomp = init_callback2(global_precomp, worker_precomp)
- infinite loop:
- wait for job on network, receive input
- output = job_callback(global_precomp, worker_precomp, input)
- send output on network
- exit loop on shutdown signal
parallel_run(global_precomp, input_array, restart file, callbacks):
- check if we're running as an MPI program
- send global_precomp to all nodes (if MPI)
- if(restart file given and exists) read restart file
- else create new restart file
- until(all jobs finished):
- if MPI:
- send next job & input to appropriate node
- if all jobs are in work, reassign unfinished ones (up to limit)
- collect outputs
- if no MPI:
- worker_precomp = init_callback1
- worker_precomp = init_callback2(global_precomp, worker_precomp)
- for(j in jobs)
- output(j) = job_callback(global_precomp, worker_precomp, input(j))
- delete restart file
- return array of outputs
parallel_destroy():
- free everything
have a context? probably yes: parallel_context
plan:
- make interface
- implement no-MPI part
- restructure singular_values.c to use interface
- implement MPI part
*/
#include <mpi.h>
#include <time.h>
typedef void (*parallel_callback_destroy)(const void*, void*);
typedef int (*parallel_callback_init)(const void*,void*);
typedef int (*parallel_callback_job)(const void*,void*,const void*,void*);
typedef struct {
int mpi_mode;
struct timespec starttime;
char processor_name[MPI_MAX_PROCESSOR_NAME];
int processor_name_len;
int rank;
int size;
MPI_Datatype order_datatype;
MPI_Datatype result_datatype;
parallel_callback_init init;
parallel_callback_job job;
parallel_callback_destroy destroy;
void *global_data;
void *node_data;
int global_data_size;
int node_data_size;
int input_size;
int output_size;
} parallel_context;
parallel_context *parallel_init();
void parallel_set_datasize_and_callbacks(parallel_context *ctx, parallel_callback_init init, parallel_callback_job job, parallel_callback_destroy destroy, int global_data_size, int node_data_size, int input_size, int output_size);
int parallel_work(parallel_context *ctx);
int parallel_run(parallel_context *ctx, const void *global_data, const void *input_array, void *output_array, unsigned int njobs, const char *restart_filename);
void parallel_destroy(parallel_context* ctx);
int mpi_rank();
void start_timer();
double runtime();
#endif