2005-03-14 23:57:21 +03:00
|
|
|
/*
|
2007-03-17 02:11:45 +03:00
|
|
|
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
|
2005-11-05 22:57:48 +03:00
|
|
|
* University Research and Technology
|
|
|
|
|
* Corporation. All rights reserved.
|
|
|
|
|
* Copyright (c) 2004-2005 The University of Tennessee and The University
|
|
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
|
* reserved.
|
2006-02-07 06:32:36 +03:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
2005-03-14 23:57:21 +03:00
|
|
|
* University of Stuttgart. All rights reserved.
|
2005-03-24 15:43:37 +03:00
|
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
|
|
|
* All rights reserved.
|
2005-03-14 23:57:21 +03:00
|
|
|
* $COPYRIGHT$
|
2006-02-07 06:32:36 +03:00
|
|
|
*
|
2005-03-14 23:57:21 +03:00
|
|
|
* Additional copyrights may follow
|
2006-02-07 06:32:36 +03:00
|
|
|
*
|
2005-03-14 23:57:21 +03:00
|
|
|
* $HEADER$
|
|
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* @file
|
|
|
|
|
*
|
|
|
|
|
* Runtime Messaging Layer (RML) Communication Interface
|
2005-03-14 23:57:21 +03:00
|
|
|
*
|
2007-07-20 05:34:02 +04:00
|
|
|
* The Runtime Messaging Layer (RML) provices basic point-to-point
|
|
|
|
|
* communication between ORTE processes. The system is available for
|
|
|
|
|
* most architectures, with some exceptions (the Cray XT3/XT4, for example).
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
#ifndef ORTE_MCA_RML_RML_H_
|
|
|
|
|
#define ORTE_MCA_RML_RML_H_
|
2005-12-01 21:28:20 +03:00
|
|
|
|
|
|
|
|
#ifdef HAVE_UNISTD_H
|
|
|
|
|
#include <unistd.h>
|
|
|
|
|
#endif
|
|
|
|
|
|
2006-02-12 04:33:29 +03:00
|
|
|
#include "opal/mca/mca.h"
|
|
|
|
|
#include "orte/mca/rml/rml_types.h"
|
2005-03-14 23:57:21 +03:00
|
|
|
|
2007-03-17 02:11:45 +03:00
|
|
|
#include "opal/mca/crs/crs.h"
|
|
|
|
|
#include "opal/mca/crs/base/base.h"
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
BEGIN_C_DECLS
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/* ******************************************************************** */
|
2006-02-07 06:32:36 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
struct orte_buffer_t;
|
|
|
|
|
struct orte_process_name_t;
|
|
|
|
|
struct orte_rml_module_t;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* ******************************************************************** */
|
|
|
|
|
|
2006-02-07 06:32:36 +03:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* RML component initialization
|
|
|
|
|
*
|
|
|
|
|
* Create an instance (module) of the given RML component. Upon
|
|
|
|
|
* returning, the module data structure should be fully populated and
|
|
|
|
|
* all functions should be usable. Non-blocking receive calls may be
|
|
|
|
|
* posted upon return from this function, although communication need
|
|
|
|
|
* not be enabled until enable_comm() call is called on the module.
|
|
|
|
|
*
|
|
|
|
|
* @return Exactly one module created by the call to the component's
|
|
|
|
|
* initialization function should be returned. The module structure
|
|
|
|
|
* should be fully populated, and the priority should be set to a
|
|
|
|
|
* reasonable value.
|
|
|
|
|
*
|
|
|
|
|
* @param[out] priority Selection priority for the given component
|
|
|
|
|
*
|
|
|
|
|
* @retval NULL An error occurred and initialization did not occur
|
|
|
|
|
* @retval non-NULL The module was successfully initialized
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef struct orte_rml_module_t* (*orte_rml_component_init_fn_t)(int *priority);
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* RML component interface
|
|
|
|
|
*
|
|
|
|
|
* Component interface for the RML framework. A public instance of
|
|
|
|
|
* this structure, called mca_rml_[component name]_component, must
|
|
|
|
|
* exist in any RML component.
|
|
|
|
|
*/
|
|
|
|
|
struct orte_rml_component_1_0_0_t {
|
|
|
|
|
/* Base component description */
|
|
|
|
|
mca_base_component_t rml_version;
|
|
|
|
|
/* Base component data block */
|
|
|
|
|
mca_base_component_data_1_0_0_t rml_data;
|
|
|
|
|
/* Component intialization function */
|
|
|
|
|
orte_rml_component_init_fn_t rml_init;
|
|
|
|
|
};
|
|
|
|
|
/** Convienence typedef */
|
|
|
|
|
typedef struct orte_rml_component_1_0_0_t orte_rml_component_t;
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/* ******************************************************************** */
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Funtion prototype for callback from non-blocking iovec send and receive
|
|
|
|
|
*
|
|
|
|
|
* Funtion prototype for callback from non-blocking iovec send and
|
|
|
|
|
* receive. The iovec pointer will be the same pointer passed to
|
|
|
|
|
* either send_nb and recv_nb. The peer memory location may not be
|
|
|
|
|
* the same, and is owned by the RML, not the calling process.
|
|
|
|
|
*
|
|
|
|
|
* @note The parameter in/out parameters are relative to the user's callback
|
|
|
|
|
* function.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] status Completion status - equivalent to the return value
|
|
|
|
|
* from blocking send/recv
|
|
|
|
|
* @param[in] peer Opaque name of peer process
|
|
|
|
|
* @param[in] msg Array of iovecs describing user buffers and lengths
|
|
|
|
|
* @param[in] count Number of elements in iovec array
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] cbdata User data passed to send_nb() or recv_nb()
|
|
|
|
|
*/
|
|
|
|
|
typedef void (*orte_rml_callback_fn_t)(int status,
|
|
|
|
|
struct orte_process_name_t* peer,
|
|
|
|
|
struct iovec* msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
void* cbdata);
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Funtion prototype for callback from non-blocking buffer send and receive
|
|
|
|
|
*
|
|
|
|
|
* Funtion prototype for callback from non-blocking buffer send and
|
|
|
|
|
* receive. The buffer may not be the same pointer passed to either
|
|
|
|
|
* send_buffer_nb and recv_buffer_nb. The peer memory location may
|
|
|
|
|
* not be the same, and is owned by the RML, not the calling process.
|
|
|
|
|
*
|
|
|
|
|
* @note The parameter in/out parameters are relative to the user's callback
|
|
|
|
|
* function.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] status Completion status - equivalent to the return value
|
|
|
|
|
* from blocking send/recv
|
|
|
|
|
* @param[in] peer Name of peer process
|
|
|
|
|
* @param[in] buffer Message buffer
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] cbdata User data passed to send_nb() or recv_nb()
|
|
|
|
|
*/
|
|
|
|
|
typedef void (*orte_rml_buffer_callback_fn_t)(int status,
|
|
|
|
|
struct orte_process_name_t* peer,
|
|
|
|
|
struct orte_buffer_t* buffer,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
void* cbdata);
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Function prototype for exception callback
|
|
|
|
|
*
|
|
|
|
|
* Function prototype for callback triggered when a communication error is detected.
|
|
|
|
|
*
|
|
|
|
|
* @note The parameter in/out parameters are relative to the user's callback
|
|
|
|
|
* function.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Name of peer process
|
|
|
|
|
* @param[in] exception Description of the error causing the exception
|
|
|
|
|
*/
|
|
|
|
|
typedef void (*orte_rml_exception_callback_t)(const orte_process_name_t* peer,
|
|
|
|
|
orte_rml_exception_t exception);
|
2006-02-07 06:32:36 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/* ******************************************************************** */
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Enable communication using the RML module
|
|
|
|
|
*
|
|
|
|
|
* Enable communication using the RML module. Before this call, only
|
|
|
|
|
* the non-blocking receive and ping interfaces may be used. After
|
|
|
|
|
* this call returns, the module must be fully functional, capable of
|
|
|
|
|
* sending and receiving data. This function will be called after the
|
|
|
|
|
* process has been assigned a proces identifier.
|
|
|
|
|
*
|
|
|
|
|
* @note While the ping interface may be used between the call to the
|
|
|
|
|
* component's initialization function and this call, care must be
|
|
|
|
|
* taken when doing so. The remote process must have already called
|
|
|
|
|
* enable_comm() or the remote process will not reply to the ping.
|
|
|
|
|
* As the ping interface is generally used by MPI processes to find a
|
|
|
|
|
* daemon to contact, this should not be a major limitation.
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS Communications successfully enabled
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
|
|
|
|
*/
|
|
|
|
|
typedef int (*orte_rml_module_enable_comm_fn_t)(void);
|
2006-02-07 06:32:36 +03:00
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Finalize the RML module
|
|
|
|
|
*
|
|
|
|
|
* Finalize the RML module, ending all communication and cleaning up
|
|
|
|
|
* all resources associated with the module. After the finalize
|
|
|
|
|
* function is called, all interface functions (and the module
|
|
|
|
|
* structure itself) are not available for use.
|
|
|
|
|
*
|
|
|
|
|
* @note Whether or not the finalize function returns successfully,
|
|
|
|
|
* the module should not be used once this function is called.
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS Success
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
|
|
|
|
*/
|
|
|
|
|
typedef int (*orte_rml_module_finalize_fn_t)(void);
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Get a "contact info" string for the local process
|
|
|
|
|
*
|
|
|
|
|
* Get a "contact info" string that can be used by other processes to
|
|
|
|
|
* share the contact information for the given process. The "contact
|
|
|
|
|
* info" string includes the process identifier for the given process
|
|
|
|
|
* and uses only basic ascii characters. It should be quoted when
|
|
|
|
|
* evaluated by a shell, although no special escaping is necessary.
|
|
|
|
|
*
|
|
|
|
|
* @note The function may return a contact info string which contains
|
|
|
|
|
* multiple addresses.
|
|
|
|
|
*
|
|
|
|
|
* @retval non-NULL The contact information for this process
|
|
|
|
|
* @retval NULL An error occurred when trying to get the current
|
|
|
|
|
* process contact info
|
|
|
|
|
*/
|
|
|
|
|
typedef char* (*orte_rml_module_get_contact_info_fn_t)(void);
|
2006-02-07 06:32:36 +03:00
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Update the RML with a remote process's contact info
|
|
|
|
|
*
|
|
|
|
|
* Update the RML with a remote process's contact information, as
|
|
|
|
|
* returned from the get_contact_info() function on the remote
|
|
|
|
|
* process. Before a send can be initiated to a remote process,
|
|
|
|
|
* either this function must be called for that process or that
|
|
|
|
|
* process must have already established a connection to the local
|
|
|
|
|
* process.
|
|
|
|
|
*
|
|
|
|
|
* @note The user may not always explicitly call this function
|
|
|
|
|
* directly, but may instead cause it to be called through one of the
|
|
|
|
|
* contact setup functions available in
|
|
|
|
|
* orte/mca/rml/base/rml_contact.h.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] contact_info The contact information string of a peer
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The contact information was successfully updated
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred during the update
|
|
|
|
|
*/
|
|
|
|
|
typedef int (*orte_rml_module_set_contact_info_fn_t)(const char *contact_info);
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Request a new name from the HNP, if one is not already assigned
|
|
|
|
|
*
|
|
|
|
|
* Request a new name from the HNP, if one is not already assigned.
|
|
|
|
|
* This function should be avoided at all costs, but is unavoidable in
|
|
|
|
|
* some instances (like when trying to get a name from a singleton or
|
|
|
|
|
* when trying to contact a persistent daemon from an orte tool.
|
|
|
|
|
*
|
|
|
|
|
* @param[out] name The new name of the process
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS A name was successfully acquired
|
|
|
|
|
* @retcal ORTE_ERR_NOT_SUPPORTED A name is already assigned to the
|
|
|
|
|
* current process
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_get_new_name_fn_t)(orte_process_name_t *name);
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* "Ping" another process to determine availability
|
|
|
|
|
*
|
|
|
|
|
* Ping another process to determine if it is available. This
|
|
|
|
|
* function only verifies that the process is alive and will allow a
|
|
|
|
|
* connection to the local process. It does *not* qualify as
|
|
|
|
|
* establishing communication with the remote process, as required by
|
|
|
|
|
* the note for set_contact_info().
|
|
|
|
|
*
|
|
|
|
|
* @param[in] contact_info The contact info string for the remote process
|
|
|
|
|
* @param[in] tv Timeout after which the ping should be failed
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCESS The process is available and will allow connections
|
|
|
|
|
* from the local process
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred during the update
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_ping_fn_t)(const char* contact_info,
|
|
|
|
|
const struct timeval* tv);
|
2005-03-14 23:57:21 +03:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/**
|
|
|
|
|
* Send an iovec blocking message
|
|
|
|
|
*
|
|
|
|
|
* Send an iovec blocking message to the specified peer. The call
|
|
|
|
|
* will return when the buffer may be modified. The return of a call
|
|
|
|
|
* to send() does not give any indication of remote completion.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Name of receiving process
|
|
|
|
|
* @param[in] msg iovec array describing send buffer
|
|
|
|
|
* @param[in] count Length of iovec array
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags Currently unused.
|
|
|
|
|
*
|
|
|
|
|
* @retval >0 Number of bytes successfully sent (will be
|
|
|
|
|
* length of all iovecs)
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_ADDRESSEE_UNKNOWN Contact information for the
|
|
|
|
|
* receiving process is not available
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
|
|
|
|
*/
|
|
|
|
|
typedef int (*orte_rml_module_send_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
struct iovec *msg,
|
|
|
|
|
int count,
|
|
|
|
|
int tag,
|
|
|
|
|
int flags);
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/**
|
|
|
|
|
* Send a buffer blocking message
|
|
|
|
|
*
|
|
|
|
|
* Send a buffer blocking message to the specified peer. The call
|
|
|
|
|
* will return when the buffer may be modified. The return of a call
|
|
|
|
|
* to send() does not give any indication of remote completion.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Name of receiving process
|
|
|
|
|
* @param[in] buffer send buffer
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags Currently unused.
|
|
|
|
|
*
|
|
|
|
|
* @retval >0 Number of bytes successfully sent (will be
|
|
|
|
|
* length of buffer)
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_ADDRESSEE_UNKNOWN Contact information for the
|
|
|
|
|
* receiving process is not available
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_send_buffer_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
struct orte_buffer_t* buffer,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags);
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Send an iovec non-blocking message
|
|
|
|
|
*
|
|
|
|
|
* Send an iovec blocking message to the specified peer. The call
|
|
|
|
|
* will return immediately, although the iovec may not be modified
|
|
|
|
|
* until the completion callback is triggered. The iovec *may* be
|
|
|
|
|
* passed to another call to send_nb before the completion callback is
|
|
|
|
|
* triggered. The callback being triggered does not give any
|
|
|
|
|
* indication of remote completion.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Name of receiving process
|
|
|
|
|
* @param[in] msg iovec array describing send buffer
|
|
|
|
|
* @param[in] count Length of iovec array
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags Currently unused
|
|
|
|
|
* @param[in] cbfunc Callback function on message comlpetion
|
|
|
|
|
* @param[in] cbdata User data to provide during completion callback
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The message was successfully started
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_ADDRESSEE_UNKNOWN Contact information for the
|
|
|
|
|
* receiving process is not available
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_send_nb_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
struct iovec* msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Send an buffer non-blocking message
|
|
|
|
|
*
|
|
|
|
|
* Send an buffer blocking message to the specified peer. The call
|
|
|
|
|
* will return immediately, although the buffer may not be modified
|
|
|
|
|
* until the completion callback is triggered. The buffer *may* be
|
|
|
|
|
* passed to another call to send_nb before the completion callback is
|
|
|
|
|
* triggered. The callback being triggered does not give any
|
|
|
|
|
* indication of remote completion.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Name of receiving process
|
|
|
|
|
* @param[in] buffer Buffer array describing send buffer
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags Currently unused
|
|
|
|
|
* @param[in] cbfunc Callback function on message comlpetion
|
|
|
|
|
* @param[in] cbdata User data to provide during completion callback
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The message was successfully started
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_ADDRESSEE_UNKNOWN Contact information for the
|
|
|
|
|
* receiving process is not available
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_send_buffer_nb_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
struct orte_buffer_t* buffer,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_buffer_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
2005-03-14 23:57:21 +03:00
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/**
|
|
|
|
|
* Receive an iovec blocking message
|
|
|
|
|
*
|
|
|
|
|
* Receive a message into a user-provided iovec. The call will not
|
|
|
|
|
* return until the buffer has been received. The remote process does
|
|
|
|
|
* not need to be connected to the current process to post the
|
|
|
|
|
* receive, so it is possible to post a receive with no sending
|
|
|
|
|
* process (ie, it will never complete). The buffer must not be
|
|
|
|
|
* currently in use with any other RML communication function during a
|
|
|
|
|
* receive call.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Peer process or ORTE_NAME_WILDCARD for wildcard receive
|
|
|
|
|
* @param[out] msg iovec array of receive buffer space
|
|
|
|
|
* @param[in] count Number of iovec entries in the array
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags May be ORTE_RML_PEEK to return up to the number
|
|
|
|
|
* of bytes provided in the iovec array without
|
|
|
|
|
* removing the message from the queue.
|
|
|
|
|
*
|
|
|
|
|
* @retval >0 Number of bytes successfully received (may be smaller
|
|
|
|
|
* than the posted buffer size
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
2005-10-06 23:39:20 +04:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_recv_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
struct iovec *msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags);
|
2006-02-07 06:32:36 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/**
|
|
|
|
|
* Receive a buffer blocking message
|
|
|
|
|
*
|
|
|
|
|
* Receive a message into a user-provided buffer. The call will not
|
|
|
|
|
* return until the buffer has been received. The remote process does
|
|
|
|
|
* not need to be connected to the current process to post the
|
|
|
|
|
* receive, so it is possible to post a receive with no sending
|
|
|
|
|
* process (ie, it will never complete). The buffer must not be
|
|
|
|
|
* currently in use with any other RML communication function during a
|
|
|
|
|
* receive call.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Peer process or ORTE_NAME_WILDCARD for wildcard receive
|
|
|
|
|
* @param[out] buffer A dss buffer to update with the received data
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags Flags modifying receive behavior
|
|
|
|
|
*
|
|
|
|
|
* @retval >0 Number of bytes successfully received (may be smaller
|
|
|
|
|
* than the posted buffer size
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
|
|
|
|
*/
|
|
|
|
|
typedef int (*orte_rml_module_recv_buffer_fn_t) (struct orte_process_name_t* peer,
|
|
|
|
|
struct orte_buffer_t *buf,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags);
|
|
|
|
|
|
2006-02-07 06:32:36 +03:00
|
|
|
|
2005-10-06 23:39:20 +04:00
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Receive an iovec non-blocking message
|
|
|
|
|
*
|
|
|
|
|
* Receive a message into a user-provided iovec. The call will not
|
|
|
|
|
* return until the buffer has been received. The remote process does
|
|
|
|
|
* not need to be connected to the current process to post the
|
|
|
|
|
* receive, so it is possible to post a receive with no sending
|
|
|
|
|
* process (ie, it will never complete). The buffer must not be
|
|
|
|
|
* currently in use with any other RML communication function during a
|
|
|
|
|
* receive call.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Peer process or ORTE_NAME_WILDCARD for wildcard receive
|
|
|
|
|
* @param[out] msg iovec array of receive buffer space
|
|
|
|
|
* @param[in] count Number of iovec entries in the array
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags May be ORTE_RML_PEEK to return up to the number
|
|
|
|
|
* of bytes provided in the iovec array without
|
|
|
|
|
* removing the message from the queue.
|
|
|
|
|
* @param[in] cbfunc Callback function on message comlpetion
|
|
|
|
|
* @param[in] cbdata User data to provide during completion callback
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The message was successfully started
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_ADDRESSEE_UNKNOWN Contact information for the
|
|
|
|
|
* receiving process is not available
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
2005-10-06 23:39:20 +04:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_recv_nb_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
struct iovec* msg,
|
|
|
|
|
int count,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
2006-02-07 06:32:36 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/**
|
|
|
|
|
* Receive a buffer non-blocking message
|
|
|
|
|
*
|
|
|
|
|
* Receive a message into a user-provided buffer. The call will not
|
|
|
|
|
* return until the buffer has been received. The remote process does
|
|
|
|
|
* not need to be connected to the current process to post the
|
|
|
|
|
* receive, so it is possible to post a receive with no sending
|
|
|
|
|
* process (ie, it will never complete). The buffer must not be
|
|
|
|
|
* currently in use with any other RML communication function during a
|
|
|
|
|
* receive call.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Peer process or ORTE_NAME_WILDCARD for wildcard receive
|
|
|
|
|
* @param[in] tag User defined tag for matching send/recv
|
|
|
|
|
* @param[in] flags May be ORTE_RML_PEEK to return up to the number
|
|
|
|
|
* of bytes provided in the iovec array without
|
|
|
|
|
* removing the message from the queue.
|
|
|
|
|
* @param[in] cbfunc Callback function on message comlpetion
|
|
|
|
|
* @param[in] cbdata User data to provide during completion callback
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The message was successfully started
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_ADDRESSEE_UNKNOWN Contact information for the
|
|
|
|
|
* receiving process is not available
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_recv_buffer_nb_fn_t)(struct orte_process_name_t* peer,
|
|
|
|
|
orte_rml_tag_t tag,
|
|
|
|
|
int flags,
|
|
|
|
|
orte_rml_buffer_callback_fn_t cbfunc,
|
|
|
|
|
void* cbdata);
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Cancel a posted non-blocking receive
|
|
|
|
|
*
|
|
|
|
|
* Attempt to cancel a posten non-blocking receive.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] peer Peer process or ORTE_NAME_WILDCARD, exactly as passed
|
|
|
|
|
* to the non-blocking receive call
|
|
|
|
|
* @param[in] tag Posted receive tag
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The receive was successfully cancelled
|
|
|
|
|
* @retval ORTE_ERR_BAD_PARAM One of the parameters was invalid
|
|
|
|
|
* @retval ORTE_ERR_NOT_FOUND A matching receive was not found
|
|
|
|
|
* @retval ORTE_ERROR An unspecified error occurred
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_recv_cancel_fn_t)(orte_process_name_t* peer,
|
|
|
|
|
orte_rml_tag_t tag);
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Register or deregister an exception callback function
|
|
|
|
|
*
|
|
|
|
|
* Register or deregister a callback when an asynchronous
|
|
|
|
|
* communication exception occurs.
|
|
|
|
|
*
|
|
|
|
|
* @param[in] cbfunc User callback
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The operation completed successfully
|
|
|
|
|
* @retval ORTE_ERROR An unspecifed error occurred
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
2007-07-20 05:34:02 +04:00
|
|
|
typedef int (*orte_rml_module_exception_fn_t)(orte_rml_exception_callback_t cbfunc);
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
|
2007-03-17 02:11:45 +03:00
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* Handle fault tolerance updates
|
|
|
|
|
*
|
|
|
|
|
* Handle fault tolerance updates
|
|
|
|
|
*
|
|
|
|
|
* @param[in] state Fault tolerance state update
|
|
|
|
|
*
|
|
|
|
|
* @retval ORTE_SUCCESS The operation completed successfully
|
|
|
|
|
* @retval ORTE_ERROR An unspecifed error occurred
|
2007-03-17 02:11:45 +03:00
|
|
|
*/
|
|
|
|
|
typedef int (*orte_rml_module_ft_event_fn_t)(int state);
|
|
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/* ******************************************************************** */
|
|
|
|
|
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
/**
|
2007-07-20 05:34:02 +04:00
|
|
|
* RML module interface
|
|
|
|
|
*
|
|
|
|
|
* Module interface to the RML communication system. A global
|
|
|
|
|
* instance of this module, orte_rml, provices an interface into the
|
|
|
|
|
* active RML interface.
|
2005-03-14 23:57:21 +03:00
|
|
|
*/
|
|
|
|
|
struct orte_rml_module_t {
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Enable communication once a process name has been assigned */
|
|
|
|
|
orte_rml_module_enable_comm_fn_t enable_comm;
|
|
|
|
|
/** Shutdown the communication system and clean up resources */
|
|
|
|
|
orte_rml_module_finalize_fn_t finalize;
|
|
|
|
|
|
|
|
|
|
/** Get contact information for local process */
|
|
|
|
|
orte_rml_module_get_contact_info_fn_t get_contact_info;
|
|
|
|
|
/** Set contact information for remote process */
|
|
|
|
|
orte_rml_module_set_contact_info_fn_t set_contact_info;
|
|
|
|
|
|
|
|
|
|
/** Get a name from the remote process */
|
|
|
|
|
orte_rml_module_get_new_name_fn_t get_new_name;
|
|
|
|
|
/** Ping process for connectivity check */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_ping_fn_t ping;
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/** Send blocking iovec message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_send_fn_t send;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Send blocking buffer message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_send_nb_fn_t send_nb;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Send non-blocking iovec message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_send_buffer_fn_t send_buffer;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Send non-blocking buffer message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_send_buffer_nb_fn_t send_buffer_nb;
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/** Receive blocking iovec message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_recv_fn_t recv;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Receive blocking buffer message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_recv_nb_fn_t recv_nb;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Receive non-blocking iovec message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_recv_buffer_fn_t recv_buffer;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Receive non-blocking buffer message */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_recv_buffer_nb_fn_t recv_buffer_nb;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Cancel posted non-blocking receive */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_recv_cancel_fn_t recv_cancel;
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/** Add callback for communication exception */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_exception_fn_t add_exception_handler;
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Delete callback for communication exception */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_exception_fn_t del_exception_handler;
|
2007-07-20 05:34:02 +04:00
|
|
|
|
|
|
|
|
/** Fault tolerance handler */
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 22:31:28 +04:00
|
|
|
orte_rml_module_ft_event_fn_t ft_event;
|
2005-03-14 23:57:21 +03:00
|
|
|
};
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Convienence typedef */
|
2005-03-14 23:57:21 +03:00
|
|
|
typedef struct orte_rml_module_t orte_rml_module_t;
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Interface for RML communication */
|
|
|
|
|
ORTE_DECLSPEC extern orte_rml_module_t orte_rml;
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/* ******************************************************************** */
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/** Macro for use in components that are of type rml v1.0.0 */
|
2005-03-14 23:57:21 +03:00
|
|
|
#define ORTE_RML_BASE_VERSION_1_0_0 \
|
|
|
|
|
/* rml v1.0 is chained to MCA v1.0 */ \
|
|
|
|
|
MCA_BASE_VERSION_1_0_0, \
|
|
|
|
|
/* rml v1.0 */ \
|
|
|
|
|
"rml", 1, 0, 0
|
|
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
/* ******************************************************************** */
|
2005-03-14 23:57:21 +03:00
|
|
|
|
|
|
|
|
|
2007-07-20 05:34:02 +04:00
|
|
|
END_C_DECLS
|
|
|
|
|
|
2005-03-14 23:57:21 +03:00
|
|
|
#endif
|