This class virtualize the cluster of PC as a set of processes that communicate. More...
This class virtualize the cluster of PC as a set of processes that communicate.
At the moment it is an MPI-like interface, with a more type aware, and simple, interface. It also give some more complex communication functionalities like Dynamic Sparse Data Exchange
Actually VCluster expose a Computation driven parallelism (MPI-like), with a plan of extending to communication driven parallelism
Definition at line 103 of file VCluster_base.hpp.
#include <VCluster_base.hpp>
Public Member Functions | |
Vcluster_base (int *argc, char ***argv) | |
Virtual cluster constructor. More... | |
MPI_Comm | getMPIComm () |
Get the MPI_Communicator (or processor group) this VCluster is using. More... | |
size_t | getProcessingUnits () |
Get the total number of processors. More... | |
size_t | size () |
Get the total number of processors. More... | |
size_t | getProcessUnitID () |
Get the process unit id. More... | |
size_t | rank () |
Get the process unit id. More... | |
template<typename T > | |
void | sum (T &num) |
Sum the numbers across all processors and get the result. More... | |
template<typename T > | |
void | max (T &num) |
Get the maximum number across all processors (or reduction with infinity norm) More... | |
template<typename T > | |
void | min (T &num) |
Get the minimum number across all processors (or reduction with insinity norm) More... | |
template<typename T > | |
void | sendrecvMultipleMessagesNBX (openfpm::vector< size_t > &prc, openfpm::vector< T > &data, openfpm::vector< size_t > prc_recv, openfpm::vector< size_t > &recv_sz, void *(*msg_alloc)(size_t, size_t, size_t, size_t, size_t, void *), void *ptr_arg, long int opt=NONE) |
Send and receive multiple messages. More... | |
template<typename T > | |
void | sendrecvMultipleMessagesNBX (openfpm::vector< size_t > &prc, openfpm::vector< T > &data, void *(*msg_alloc)(size_t, size_t, size_t, size_t, size_t, void *), void *ptr_arg, long int opt=NONE) |
Send and receive multiple messages. More... | |
void | sendrecvMultipleMessagesNBX (size_t n_send, size_t sz[], size_t prc[], void *ptr[], size_t n_recv, size_t prc_recv[], size_t sz_recv[], void *(*msg_alloc)(size_t, size_t, size_t, size_t, size_t, void *), void *ptr_arg, long int opt=NONE) |
Send and receive multiple messages. More... | |
void | sendrecvMultipleMessagesNBX (size_t n_send, size_t sz[], size_t prc[], void *ptr[], size_t n_recv, size_t prc_recv[], void *(*msg_alloc)(size_t, size_t, size_t, size_t, size_t, void *), void *ptr_arg, long int opt=NONE) |
Send and receive multiple messages. More... | |
void | sendrecvMultipleMessagesNBX (size_t n_send, size_t sz[], size_t prc[], void *ptr[], void *(*msg_alloc)(size_t, size_t, size_t, size_t, size_t, void *), void *ptr_arg, long int opt=NONE) |
Send and receive multiple messages. More... | |
bool | send (size_t proc, size_t tag, const void *mem, size_t sz) |
Send data to a processor. More... | |
template<typename T , typename Mem , typename gr > | |
bool | send (size_t proc, size_t tag, openfpm::vector< T, Mem, gr > &v) |
Send data to a processor. More... | |
bool | recv (size_t proc, size_t tag, void *v, size_t sz) |
Recv data from a processor. More... | |
template<typename T , typename Mem , typename gr > | |
bool | recv (size_t proc, size_t tag, openfpm::vector< T, Mem, gr > &v) |
Recv data from a processor. More... | |
template<typename T , typename Mem , typename gr > | |
bool | allGather (T &send, openfpm::vector< T, Mem, gr > &v) |
Gather the data from all processors. More... | |
template<typename T , typename Mem , typename gr > | |
bool | Bcast (openfpm::vector< T, Mem, gr > &v, size_t root) |
Broadcast the data to all processors. More... | |
void | execute () |
Execute all the requests. More... | |
void | clear () |
Release the buffer used for communication. More... | |
Data Fields | |
openfpm::vector< size_t > | sz_recv_tmp |
Protected Attributes | |
openfpm::vector< BHeapMemory > | recv_buf |
Receive buffers. | |
Private Member Functions | |
Vcluster_base & | operator= (const Vcluster_base &) |
disable operator= | |
Vcluster_base (const Vcluster_base &) | |
disable copy constructor | |
Private Attributes | |
Vcluster_log | log |
log file | |
size_t | NBX_cnt |
openfpm::vector< size_t > | proc_com |
openfpm::vector< int > | map_scatter |
vector that contain the scatter map (it is basically an array of one) | |
openfpm::vector< MPI_Request > | req |
vector of MPI requests | |
openfpm::vector< MPI_Status > | stat |
vector of MPI status | |
std::vector< int > | post_exe |
vector of functions to execute after all the request has been performed | |
int | m_size |
number of processes | |
int | m_rank |
actual rank | |
int | numPE = 1 |
number of processing unit per process | |
std::vector< red > | r |
openfpm::vector< void * > | ptr_send |
vector of pointers of send buffers | |
openfpm::vector< size_t > | sz_send |
vector of the size of send buffers | |
MPI_Request | bar_req |
barrier request | |
MPI_Status | bar_stat |
barrier status | |
|
inline |
Virtual cluster constructor.
argc | pointer to arguments counts passed to the program |
argv | pointer to arguments vector passed to the program |
Definition at line 219 of file VCluster_base.hpp.
|
inline |
Gather the data from all processors.
send a primitive data T receive the same primitive T from all the other processors
v | vector to receive (automaticaly resized) |
send | data to send |
Definition at line 1002 of file VCluster_base.hpp.
|
inline |
Broadcast the data to all processors.
broadcast a vector of primitives.
v | vector to send in the case of the root processor and vector where to receive in the case of non-root |
root | processor (who broadcast) |
Definition at line 1036 of file VCluster_base.hpp.
|
inline |
Release the buffer used for communication.
Definition at line 1073 of file VCluster_base.hpp.
|
inline |
Execute all the requests.
Definition at line 1054 of file VCluster_base.hpp.
|
inline |
Get the MPI_Communicator (or processor group) this VCluster is using.
Definition at line 322 of file VCluster_base.hpp.
|
inline |
Get the total number of processors.
Definition at line 332 of file VCluster_base.hpp.
|
inline |
|
inline |
Get the maximum number across all processors (or reduction with infinity norm)
num | to reduce |
Definition at line 402 of file VCluster_base.hpp.
|
inline |
Get the minimum number across all processors (or reduction with insinity norm)
num | to reduce |
Definition at line 422 of file VCluster_base.hpp.
|
inline |
Get the process unit id.
It is the same as getProcessUnitID()
Definition at line 370 of file VCluster_base.hpp.
|
inline |
Recv data from a processor.
proc | processor id |
tag | id |
v | buffer to send |
sz | size of the buffer |
Definition at line 943 of file VCluster_base.hpp.
|
inline |
Recv data from a processor.
proc | processor id |
tag | id |
v | vector to send |
Definition at line 973 of file VCluster_base.hpp.
|
inline |
Send data to a processor.
proc | processor id |
tag | id |
mem | buffer with the data to send |
sz | size |
Definition at line 877 of file VCluster_base.hpp.
|
inline |
Send data to a processor.
proc | processor id |
tag | id |
v | buffer to send |
Definition at line 908 of file VCluster_base.hpp.
|
inline |
Send and receive multiple messages.
It send multiple messages to a set of processors the and receive multiple messages from another set of processors, all the processor must call this function. In this particular case the receiver know from which processor is going to receive.
suppose the following situation the calling processor want to communicate
prc | list of processor with which it should communicate [1,1,6,7,8] |
data | data to send for each processors in contain a pointer to some type T this type T must have a method size() that return the size of the data-structure |
prc_recv | processor that receive data |
recv_sz | for each processor indicate the size of the data received |
msg_alloc | This is a call-back with the purpose of allocate space for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have the following 6 parameters in the call-back are in order:
|
ptr_arg | data passed to the call-back function specified |
opt | options, NONE (ignored in this moment) |
Definition at line 479 of file VCluster_base.hpp.
|
inline |
Send and receive multiple messages.
It send multiple messages to a set of processors the and receive multiple messages from another set of processors, all the processor must call this function
suppose the following situation the calling processor want to communicate
prc | list of processors you should communicate with [1,1,6,7,8] |
data | vector containing the data to send [v=vector<vector<int>>, v.size()=4, T=vector<int>], T at the moment is only tested for vectors of 0 or more generic elements (without pointers) |
msg_alloc | This is a call-back with the purpose to allocate space for the incoming messages and give back a valid pointer, supposing that this call-back has been triggered by the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have the following 6 parameters in the call-back in order:
|
ptr_arg | data passed to the call-back function specified |
opt | options, only NONE supported |
Definition at line 542 of file VCluster_base.hpp.
|
inline |
Send and receive multiple messages.
It send multiple messages to a set of processors the and receive multiple messages from another set of processors, all the processor must call this function. In this particular case the receiver know from which processor is going to receive.
suppose the following situation the calling processor want to communicate
n_send | number of send for this processor [4] |
prc | list of processor with which it should communicate [1,1,6,7,8] |
sz | the array contain the size of the message for each processor (zeros must not be presents) [100,100,50,48,70] |
ptr | array that contain the pointers to the message to send |
msg_alloc | This is a call-back with the purpose of allocate space for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have the following 6 parameters in the call-back are in order:
|
ptr_arg | data passed to the call-back function specified |
opt | options, NONE (ignored in this moment) |
Definition at line 605 of file VCluster_base.hpp.
|
inline |
Send and receive multiple messages.
It send multiple messages to a set of processors the and receive multiple messages from another set of processors, all the processor must call this function. In this particular case the receiver know from which processor is going to receive, but does not know the size.
suppose the following situation the calling processor want to communicate
n_send | number of send for this processor [4] |
prc | list of processor with which it should communicate [1,1,6,7,8] |
sz | the array contain the size of the message for each processor (zeros must not be presents) [100,100,50,48,70] |
ptr | array that contain the pointers to the message to send |
msg_alloc | This is a call-back with the purpose of allocate space for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have the following 6 parameters in the call-back are in order:
|
ptr_arg | data passed to the call-back function specified |
opt | options, NONE (ignored in this moment) |
Definition at line 673 of file VCluster_base.hpp.
|
inline |
Send and receive multiple messages.
It send multiple messages to a set of processors the and receive multiple messages from another set of processors, all the processor must call this function
suppose the following situation the calling processor want to communicate
n_send | number of send for this processor [4] |
prc | list of processor with which it should communicate [1,1,6,7,8] |
sz | the array contain the size of the message for each processor (zeros must not be presents) [100,100,50,48,70] |
ptr | array that contain the pointers to the message to send |
msg_alloc | This is a call-back with the purpose of allocate space for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have the following 6 parameters in the call-back are in order:
|
ptr_arg | data passed to the call-back function specified |
opt | options, NONE (ignored in this moment) |
Definition at line 751 of file VCluster_base.hpp.
|
inline |
Get the total number of processors.
It is the same as getProcessingUnits()
Definition at line 346 of file VCluster_base.hpp.
|
inline |
Sum the numbers across all processors and get the result.
num | to reduce, input and output |
Definition at line 382 of file VCluster_base.hpp.
|
private |
NBX has a potential pitfall that must be addressed, NBX Send all the messages and probe for incoming messages, if there is an incoming message it receive it producing an acknowledge notification on the sending processor. When all the sends has been acknowledged, the processor call the MPI_Ibarrier when all the processors call MPI_Ibarrier all send has been received. While the processors are waiting for the MPI_Ibarrier to complete, all processors are still probing for incoming message, Unfortunately some processor can quit the MPI_Ibarrier before others and this mean that some processor can exit the probing status before others, these processors can in theory start new communications while the other processor are still in probing status producing a wrong send/recv association to resolve this problem an incremental NBX_cnt is used as message TAG to distinguish that the messages come from other send or subsequent NBX procedures
Definition at line 122 of file VCluster_base.hpp.
|
private |
temporal vector used for meta-communication ( or meta-data before the real communication )
Definition at line 126 of file VCluster_base.hpp.
|
private |
This buffer is a temporal buffer for reductions
MPI_Iallreduce does not accept recv and send buffer to be the same r is used to overcome this problem (is given as second parameter) after the execution the data is copied back
Definition at line 160 of file VCluster_base.hpp.