dlb_talp.h¶
- Date:
Mon Jul 15 2024
NAME¶
dlb_talp.h
SYNOPSIS¶
#include <time.h>#include <stdint.h>Data Structures¶
Macros¶
Enumerations¶
Functions¶
Macro Definition Documentation¶
#define DLB_MPI_REGION NULL¶
#define DLB_LAST_OPEN_REGION (void*)1¶
Enumeration Type Documentation¶
anonymous enum¶
Enumerator
DLB_MONITOR_NAME_MAX
Function Documentation¶
int DLB_TALP_Attach (void)¶
Attach current process to DLB system as TALP administrator.
Returns
DLB_SUCCESS on success
Once the process is attached to DLB as TALP administrator, it may perform the below actions described in this file. This way, the process is able to obtain some TALP values such as time spent in computation or MPI for each of the DLB running processes.
int DLB_TALP_Detach (void)¶
Detach current process from DLB system.
Returns
DLB_SUCCESS on success
DLB_ERR_NOSHMEM if cannot find shared memory to detach from
If previously attached, a process must call this function to correctly close file descriptors and clean data.
int DLB_TALP_GetNumCPUs (int * ncpus)¶
Get the number of CPUs in the node.
Parameters
ncpus the number of CPUs
Returns
DLB_SUCCESS on success
int DLB_TALP_GetPidList (int * pidlist, int * nelems, int max_len)¶
Get the list of running processes registered in the DLB system.
Parameters
pidlist The output listnelems Number of elements in the listmax_len Max capacity of the list
Returns
DLB_SUCCESS on success
DLB_ERR_NOSHMEM if cannot find shared memory
int DLB_TALP_GetTimes (int pid, double * mpi_time, double * useful_time)¶
Get the CPU time spent on MPI and useful computation for the given process.
Parameters
pid target Process IDmpi_time CPU time spent on MPI in secondsuseful_time CPU time spend on useful computation in seconds
Returns
DLB_SUCCESS on success
DLB_ERR_NOPROC if target pid is not registered in the DLB system
int DLB_TALP_GetNodeTimes (const char * name, dlb_node_times_t* node_times_list, int * nelems, int max_len)¶
Get the list of raw times for the specified region.
Parameters
name Name to identify the regionnode_times_list The output listnelems Number of elements in the listmax_len Max capacity of the list
Returns
DLB_SUCCESS on success
DLB_ERR_NOSHMEM if cannot find shared memory
int DLB_TALP_QueryPOPNodeMetrics (const char * name, dlb_node_metrics_t* node_metrics)¶
From either 1st or 3rd party, query node metrics for one region.
Parameters
name Name to identify the regionnode_metrics Allocated structure where the collected metrics will be stored
Returns
DLB_SUCCESS on success
DLB_ERR_NOENT if no data for the given name
If called from a third party, this function requires the application to run with DLB_ARGS+=’ –talp-external-profiler’
const dlb_monitor_t* DLB_MonitoringRegionGetMPIRegion (void)¶
Get the pointer of the implicit MPI Monitoring Region. DISCLAIMER: The functions declared above are intended to be called from 1st-party or 3rd-party programs indistinctly; that is, DLB applications, or external profilers as long as they invoke DLB_TALP_Attach.
The functions declared below are intended to be called only from 1st-party programs, and they should return an error if they are called from external profilers.
This header file may be split in two in the next major release.
Returns
monitor handle to be used on queries, or NULL if TALP is not enabled
dlb_monitor_t* DLB_MonitoringRegionRegister (const char * name)¶
Register a new Monitoring Region.
Parameters
name Name to identify the new region
Returns
monitor handle to be used on subsequent calls, or NULL if TALP is not enabled
int DLB_MonitoringRegionReset (dlb_monitor_t* handle)¶
Reset monitoring region.
Parameters
handle Monitoring handle that identifies the region, or DLB_MPI_REGION
Returns
DLB_SUCCESS on success
DLB_ERR_NOTALP if TALP is not enabled
int DLB_MonitoringRegionStart (dlb_monitor_t* handle)¶
Start (or unpause) monitoring region.
Parameters
handle Monitoring handle that identifies the region, or DLB_MPI_REGION
Returns
DLB_SUCCESS on success
DLB_ERR_NOTALP if TALP is not enabled
int DLB_MonitoringRegionStop (dlb_monitor_t* handle)¶
Stop (or pause) monitoring region.
Parameters
handle Monitoring handle that identifies the region, DLB_MPI_REGION, or DLB_LAST_OPEN_REGION
Returns
DLB_SUCCESS on success
DLB_ERR_NOTALP if TALP is not enabled
DLB_ERR_NOENT if DLB_LAST_OPEN_REGION does not match any region
int DLB_MonitoringRegionReport (const dlb_monitor_t* handle)¶
Print a report to stderr of the monitoring region.
Parameters
handle Monitoring handle that identifies the region, or DLB_MPI_REGION
Returns
DLB_SUCCESS on success
DLB_ERR_NOTALP if TALP is not enabled
int DLB_MonitoringRegionsUpdate (void)¶
Update all monitoring regions.
Returns
DLB_SUCCESS on success
Monitoring regions are only updated in certain situations, like when starting/stopping a region, or finalizing MPI. This routine forces the update of all started monitoring regions
int DLB_TALP_CollectPOPMetrics (dlb_monitor_t* monitor,dlb_pop_metrics_t* pop_metrics)¶
Perform an MPI collective communication to collect POP metrics.
Parameters
monitor Monitoring handle that identifies the region, or DLB_MPI_REGION macro (NULL) if implicit MPI regionpop_metrics Allocated structure where the collected metrics will be stored
Returns
DLB_SUCCESS on success
DLB_ERR_NOTALP if TALP is not enabled
int DLB_TALP_CollectPOPNodeMetrics (dlb_monitor_t* monitor,dlb_node_metrics_t* node_metrics)¶
Perform a node collective communication to collect TALP node metrics.
Parameters
monitor Monitoring handle that identifies the region, or DLB_MPI_REGION macro (NULL) if implicit MPI regionnode_metrics Allocated structure where the collected metrics will be stored
Returns
DLB_SUCCESS on success
DLB_ERR_NOTALP if TALP is not enabled
DLB_ERR_NOCOMP if support for barrier is disabled, i.e., –no-barrier
This functions performs a node barrier to collect the data. All processes that are running in the node must invoke this function.
