上一篇文章说道,初始化失败会有一个函数调用:
ompi_mpi_errors_are_fatal_comm_handler(NULL, NULL, message);
所以这里简单地进入了 ompi_mpi_errors_are_fatal_comm_handler 函数:
看到其头文件 errhandler_predefined.h :
#ifndef OMPI_ERRHANDLER_PREDEFINED_H
#define OMPI_ERRHANDLER_PREDEFINED_H
#include "ompi_config.h"
struct ompi_communicator_t;
struct ompi_file_t;
struct ompi_win_t;
/**
* Handler function for MPI_ERRORS_ARE_FATAL //---------------看到了吗?fatal_error
*/
OMPI_DECLSPEC void ompi_mpi_errors_are_fatal_comm_handler(struct ompi_communicator_t **comm,
int *error_code, ...);
OMPI_DECLSPEC void ompi_mpi_errors_are_fatal_file_handler(struct ompi_file_t **file,
int *error_code, ...);
OMPI_DECLSPEC void ompi_mpi_errors_are_fatal_win_handler(struct ompi_win_t **win,
int *error_code, ...);
/**
* Handler function for MPI_ERRORS_RETURN //---------------- error_return
*/
OMPI_DECLSPEC void ompi_mpi_errors_return_comm_handler(struct ompi_communicator_t **comm,
int *error_code, ...);
OMPI_DECLSPEC void ompi_mpi_errors_return_file_handler(struct ompi_file_t **file,
int *error_code, ...);
OMPI_DECLSPEC void ompi_mpi_errors_return_win_handler(struct ompi_win_t **win,
int *error_code, ...);
#endif /* OMPI_ERRHANDLER_PREDEFINED_H */
跳去它的实现文件 errhandler_predefined.c 中看对应函数:
void ompi_mpi_errors_are_fatal_comm_handler(struct ompi_communicator_t **comm,
int *error_code, ...)
{
char *name;
struct ompi_communicator_t *abort_comm;
va_list arglist;
va_start(arglist, error_code);
if (NULL != comm) {
name = (*comm)->c_name;
abort_comm = *comm;
} else {
name = NULL;
abort_comm = NULL;
}
backend_fatal("communicator", abort_comm, name, error_code, arglist);
va_end(arglist);
}
映入眼帘的是 ompi_communicator_t 类,MPI中的通信子应该就是通过这个类来实现的,后期要重点学习
这里涉及到不变参数列表的知识,参考此处: http://www.cnblogs.com/hanyonglu/archive/2011/05/07/2039916.html
放到此处的具体场景,也就是arglist只有message一个变量,其实也就是:
static const char FUNC_NAME[] = "MPI_Init";
这里做了一些处理之后,调用 backend_fatal 函数,这是在本文件中的一个局部函数, 它做了什么呢? :
static void backend_fatal(char *type, struct ompi_communicator_t *comm,
char *name, int *error_code,
va_list arglist)
{
/* We only want aggregation while the rte is initialized */
if (ompi_rte_initialized) {
backend_fatal_aggregate(type, comm, name, error_code, arglist);
} else {
backend_fatal_no_aggregate(type, comm, name, error_code, arglist);
}
/* In most instances the communicator will be valid. If not, we are either early in
* the initialization or we are dealing with a window. Thus, it is good enough to abort
* on MPI_COMM_SELF, the error will propagate.
*/
if (comm == NULL) {
comm = &ompi_mpi_comm_self.comm;
}
if (NULL != error_code) {
ompi_mpi_abort(comm, *error_code);
} else {
ompi_mpi_abort(comm, 1);
}
}
backend_fatal_aggregate 函数是和聚合有关的,我们先跳过。————留下疑点1
因为我们传入的 error_code 参数为NULL, 最后程序会进入 ompi_mpi_abort(comm, 1);
它在 mpiruntime.h 头文件中,实现在 ompi_mpi_abort.c 文件中,贴上它的函数定义代码:
// 目前传入的参数是: &ompi_mpi_comm_self.comm, 1
int
ompi_mpi_abort(struct ompi_communicator_t* comm,
int errcode)
{
// 1. 我们要看的第一部分代码: 防止递归,获取节点名称
char *host, hostname[OPAL_MAXHOSTNAMELEN];
pid_t pid = 0;
/* Protection for recursive invocation */
if (have_been_invoked) {
return OMPI_SUCCESS;
}
have_been_invoked = true;
/* If MPI is initialized, we know we have a runtime nodename, so
use that. Otherwise, call gethostname. */
// 疑问2: rte到底是什么?这个估计后续的深入了解会接触的更多,可能是runtime environment
if (ompi_rte_initialized) {
// host代表的也就是rank, 存储在 ompi_process_info 这个结构体中, ————疑问3
host = ompi_process_info.nodename;
} else {
gethostname(hostname, sizeof(hostname));
host = hostname;
}
pid = getpid();
// 2. 我们要看的第二部分代码: 打印函数调用堆栈
/* Should we print a stack trace? Not aggregated because they
might be different on all processes. */
if (opal_abort_print_stack) {
char **messages;
int len, i;
if (OPAL_SUCCESS == opal_backtrace_buffer(&messages, &len)) {
for (i = 0; i < len; ++i) {
fprintf(stderr, "[%s:%d] [%d] func:%s\n", host, (int) pid,
i, messages[i]);
fflush(stderr);
}
free(messages);
} else {
/* This will print an message if it's unable to print the
backtrace, so we don't need an additional "else" clause
if opal_backtrace_print() is not supported. */
opal_backtrace_print(stderr, NULL, 1);
}
}
// 3. 第三部分代码: abort之前的自旋等待
/* Should we wait for a while before aborting? */
if (0 != opal_abort_delay) {
if (opal_abort_delay < 0) {
fprintf(stderr ,"[%s:%d] Looping forever (MCA parameter opal_abort_delay is < 0)\n",
host, (int) pid);
fflush(stderr);
while (1) {
sleep(5);
}
} else {
fprintf(stderr, "[%s:%d] Delaying for %d seconds before aborting\n",
host, (int) pid, opal_abort_delay);
do {
sleep(1);
} while (--opal_abort_delay > 0);
}
}
// 4. 第四部分: RTE未初始化的情况下,看ompi_mpi_finalized的情况是哪种
/* If the RTE isn't setup yet/any more, then don't even try
killing everyone. Sorry, Charlie... */
if (!ompi_rte_initialized) {
fprintf(stderr, "[%s:%d] Local abort %s completed successfully, but am not able to aggregate error messages, and not able to guarantee that all other processes were killed!\n",
host, (int) pid, ompi_mpi_finalized ?
"after MPI_FINALIZE started" : "before MPI_INIT completed");
_exit(errcode == 0 ? 1 : errcode);
}
// 5.第五部分: 有了communicator, 干掉进程集————疑问4:那是不是说,如果有一个进程初始化失败,整个进程集都会挂掉呢?
/* If OMPI is initialized and we have a non-NULL communicator,
then try to kill just that set of processes */
if (ompi_mpi_initialized && !ompi_mpi_finalized && NULL != comm) {
try_kill_peers(comm, errcode); // 疑问 5
}
// 6. 第六部分: 很少情况会执行到这里,abort运行环境
/* We can fall through to here in a few cases:
1. The attempt to kill just a subset of peers via
try_kill_peers() failed (e.g., as of July 2014, ORTE does
returns NOT_IMPLENTED from orte_rte_abort_peers()).
2. MPI wasn't initialized, was already finalized, or we got a
NULL communicator.
In all of these cases, the only sensible thing left to do is to
kill the entire job. Wah wah. */
ompi_rte_abort(errcode, NULL); //疑问 6
/* Does not return */
}
为了章节结构好看,本篇就到此,留下6个疑问,再次重申一下:
1. backend_fatal_aggregate 函数
2: rte到底是什么?这个估计后续的深入了解会接触的更多,可能是runtime environment
3. host代表的也就是rank, 存储在 ompi_process_info 这个结构体中
4.那是不是说,如果有一个进程初始化失败,整个进程集都会挂掉呢?
5. try_kill_peers 函数
6. ompi_rte_abort 函数