| Geant4 Cross Reference |
1 //
2 // ********************************************************************
3 // * License and Disclaimer *
4 // * *
5 // * The Geant4 software is copyright of the Copyright Holders of *
6 // * the Geant4 Collaboration. It is provided under the terms and *
7 // * conditions of the Geant4 Software License, included in the file *
8 // * LICENSE and available at http://cern.ch/geant4/license . These *
9 // * include a list of copyright holders. *
10 // * *
11 // * Neither the authors of this software system, nor their employing *
12 // * institutes,nor the agencies providing financial support for this *
13 // * work make any representation or warranty, express or implied, *
14 // * regarding this software system or assume any liability for its *
15 // * use. Please see the license in the file LICENSE and URL above *
16 // * for the full disclaimer and the limitation of liability. *
17 // * *
18 // * This code implementation is the result of the scientific and *
19 // * technical work of the GEANT4 collaboration. *
20 // * By using, copying, modifying or distributing the software (or *
21 // * any work based on the software) you agree to acknowledge its *
22 // * use in resulting scientific publications, and indicate your *
23 // * acceptance of all terms of the Geant4 Software license. *
24 // ********************************************************************
25 //
26 #include "G4VUserMPIrunMerger.hh"
27
28 #include "G4MPImanager.hh"
29 #include "G4MPIutils.hh"
30
31 #include <algorithm>
32 #include <assert.h>
33 #include <functional>
34 #include <mpi.h>
35
36 G4VUserMPIrunMerger::G4VUserMPIrunMerger(const G4Run* aRun, G4int destination, G4int ver)
37 : outputBuffer(nullptr),
38 outputBufferSize(0),
39 outputBufferPosition(0),
40 ownsBuffer(false),
41 destinationRank(destination),
42 run(const_cast<G4Run*>(aRun)),
43 commSize(0),
44 verbose(ver),
45 bytesSent(0)
46 {}
47
48 #define DMSG(LVL, MSG) \
49 { \
50 if (verbose > LVL) { \
51 G4cout << MSG << G4endl; \
52 } \
53 }
54
55 void G4VUserMPIrunMerger::Send(const unsigned int destination)
56 {
57 assert(run != nullptr);
58 G4int nevts = run->GetNumberOfEvent();
59 DMSG(1, "G4VUserMPIrunMerger::Send() : Sending a G4run (" << run << ") with " << nevts
60 << " events to: " << destination);
61 input_userdata.clear();
62 Pack(); // User code
63 InputUserData(&nevts, MPI::INT, 1);
64
65 DestroyBuffer();
66 G4int newbuffsize = 0;
67 for (const const_registered_data& el : input_userdata) {
68 newbuffsize += (el.dt.Get_size() * el.count);
69 }
70 char* buffer = new char[newbuffsize];
71 // Avoid complains from valgrind (i'm not really sure why this is needed, but, beside the
72 // small cpu penalty, we can live with that).)
73 std::fill(buffer, buffer + newbuffsize, 0);
74 ownsBuffer = true;
75 SetupOutputBuffer(buffer, newbuffsize, 0);
76 DMSG(3, "Buffer size: " << newbuffsize << " bytes at: " << (void*)outputBuffer);
77
78 // Now userdata contains all data to be send, do the real packing
79 for (const const_registered_data& el : input_userdata) {
80 #ifdef G4MPI_USE_MPI_PACK_NOT_CONST
81 MPI_Pack(const_cast<void*>(el.p_data), el.count, el.dt,
82 #else
83 MPI_Pack(el.p_data, el.count, el.dt,
84 #endif
85 outputBuffer, outputBufferSize, &outputBufferPosition, COMM_G4COMMAND_);
86 }
87 assert(outputBufferSize == outputBufferPosition);
88 COMM_G4COMMAND_.Send(outputBuffer, outputBufferSize, MPI::PACKED, destination,
89 G4MPImanager::kTAG_RUN);
90 bytesSent += outputBufferSize;
91 DMSG(2, "G4VUserMPIrunMerger::Send() : Done ");
92 }
93
94 void G4VUserMPIrunMerger::Receive(const unsigned int source)
95 {
96 const MPI::Intracomm* parentComm = G4MPImanager::GetManager()->GetComm();
97 DMSG(1, "G4VUserMPIrunMerger::Receive(...) , this rank : " << parentComm->Get_rank()
98 << " and receiving from : " << source);
99 // DestroyBuffer();
100 // Receive from all but one
101 // for (G4int rank = 0; rank < commSize-1; ++rank)
102 //{
103 MPI::Status status;
104 COMM_G4COMMAND_.Probe(source, G4MPImanager::kTAG_RUN, status);
105 // const G4int source = status.Get_source();
106 const G4int newbuffsize = status.Get_count(MPI::PACKED);
107 DMSG(2, "Preparing to receive buffer of size: " << newbuffsize);
108 char* buffer = outputBuffer;
109 if (newbuffsize > outputBufferSize) {
110 DMSG(3, "New larger buffer expected, resize");
111 // New larger buffer incoming, recreate buffer
112 delete[] outputBuffer;
113 buffer = new char[newbuffsize];
114 // Avoid complains from valgrind (i'm not really sure why this is needed, but, beside the
115 // small cpu penalty, we can live with that).)
116 std::fill(buffer, buffer + newbuffsize, 0);
117 ownsBuffer = true;
118 }
119 SetupOutputBuffer(buffer, newbuffsize, 0);
120 COMM_G4COMMAND_.Recv(buffer, newbuffsize, MPI::PACKED, source, G4MPImanager::kTAG_RUN, status);
121 DMSG(3, "Buffer Size: " << outputBufferSize << " bytes at: " << (void*)outputBuffer);
122 output_userdata.clear();
123 // User code, if implemented will return the concrete G4Run class
124 G4Run* aNewRun = UnPack();
125 if (aNewRun == nullptr) aNewRun = new G4Run;
126 // Add number of events counter
127 G4int nevets = 0;
128 OutputUserData(&nevets, MPI::INT, 1);
129 // now userdata contains all data references, do the real unpacking
130 for (const registered_data& el : output_userdata) {
131 MPI_Unpack(outputBuffer, outputBufferSize, &outputBufferPosition, el.p_data, el.count, el.dt,
132 COMM_G4COMMAND_);
133 }
134 for (G4int i = 0; i < nevets; ++i)
135 aNewRun->RecordEvent(nullptr);
136
137 // Now merge received MPI run with global one
138 DMSG(2, "Before G4Run::Merge : " << run->GetNumberOfEvent());
139 run->Merge(aNewRun);
140 DMSG(2, "After G4Run::Merge : " << run->GetNumberOfEvent());
141 delete aNewRun;
142 //}
143 }
144
145 void G4VUserMPIrunMerger::Merge()
146 {
147 // G4cout << "G4VUserMPIrunMerger::Merge called" << G4endl;
148
149 DMSG(0, "G4VUserMPIrunMerger::Merge called");
150 const MPI::Intracomm* parentComm = G4MPImanager::GetManager()->GetComm();
151 const unsigned int myrank = parentComm->Get_rank();
152 commSize = G4MPImanager::GetManager()->GetActiveSize();
153 // do not include extra worker in this communication
154
155 if (commSize == 1) {
156 DMSG(1, "Comm world size is 1, nothing to do");
157 return;
158 }
159 COMM_G4COMMAND_ = parentComm->Dup();
160 bytesSent = 0;
161 const G4double sttime = MPI::Wtime();
162
163 // Use G4MPIutils to optimize communications between ranks
164 typedef std::function<void(unsigned int)> handler_t;
165 using std::placeholders::_1;
166 handler_t sender = std::bind(&G4VUserMPIrunMerger::Send, this, _1);
167 handler_t receiver = std::bind(&G4VUserMPIrunMerger::Receive, this, _1);
168 std::function<void(void)> barrier = std::bind(&MPI::Intracomm::Barrier, &COMM_G4COMMAND_);
169 // G4cout << "go to G4mpi::Merge" << G4endl;
170 G4mpi::Merge(sender, receiver, barrier, commSize, myrank);
171
172 // OLD Style p2p communications
173 /*
174 if ( myrank != destinationRank ) {
175 DMSG(0,"Comm world size: "<<commSize<<" this rank is: "
176 <<myrank<<" sending to rank "<<destinationRank);
177 Send(destinationRank);
178 } else {
179 DMSG(1,"Comm world size: "<<commSize<<" this rank is: "
180 <<myrank<<" receiving. ");
181 for ( unsigned int i = 0 ; i<commSize ; ++i) {
182 if ( i != myrank ) Receive(i);
183 }
184 }
185 */
186 const G4double elapsed = MPI::Wtime() - sttime;
187 long total = 0;
188 COMM_G4COMMAND_.Reduce(&bytesSent, &total, 1, MPI::LONG, MPI::SUM, destinationRank);
189 if (verbose > 0 && myrank == destinationRank) {
190 // Collect from ranks how much data was sent around
191 G4cout << "G4VUserMPIrunMerger::Merge() - data transfer performances: "
192 << double(total) / 1000. / elapsed << " kB/s"
193 << " (Total Data Transfer= " << double(total) / 1000 << " kB in " << elapsed << " s)."
194 << G4endl;
195 }
196
197 COMM_G4COMMAND_.Free();
198 DMSG(0, "G4VUserMPIrunMerger::Merge done");
199 }
200