jirayu_sri/stockfish
1
0
Fork 0
mirror of https://github.com/official-stockfish/Stockfish.git synced 2026-07-23 05:07:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement proper stop signalling from root node

Browse Source 
Previous behavior was to wait on all nodes to finish their search on their own TM and aggregate to root node via a blocking MPI_Allreduce call. This seems to be problematic.

In this commit a proper non-blocking signalling barrier was implemented to use TM from root node to control the cluster search, and disable TM on all non-root nodes.

Also includes some cosmetic fix to the nodes/NPS display.
This commit is contained in:
noobpwnftw
2018-12-29 15:34:55 +01:00
committed by Stéphane Nicolet
parent 3b7b632aa5
commit 8a95d269eb
6 changed files with 48 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/cluster.cpp
+30
View File
@@ -37,10 +37,13 @@ namespace Cluster {
static int world_rank = MPI_PROC_NULL;
static int world_size = 0;
static bool stop_signal = false;
static MPI_Request reqStop = MPI_REQUEST_NULL;
static MPI_Comm InputComm = MPI_COMM_NULL;
static MPI_Comm TTComm = MPI_COMM_NULL;
static MPI_Comm MoveComm = MPI_COMM_NULL;
static MPI_Comm StopComm = MPI_COMM_NULL;
static MPI_Datatype TTEntryDatatype = MPI_DATATYPE_NULL;
static std::vector<TTEntry> TTBuff;
@@ -104,6 +107,7 @@ void init() {
MPI_Comm_dup(MPI_COMM_WORLD, &InputComm);
MPI_Comm_dup(MPI_COMM_WORLD, &TTComm);
MPI_Comm_dup(MPI_COMM_WORLD, &MoveComm);
MPI_Comm_dup(MPI_COMM_WORLD, &StopComm);
}
void finalize() {
@@ -131,6 +135,32 @@ bool getline(std::istream& input, std::string& str) {
return state;
}
void sync_start() {
stop_signal = false;
// Start listening to stop signal
if (!is_root())
MPI_Ibarrier(StopComm, &reqStop);
}
void sync_stop() {
if (is_root()) {
if (!stop_signal && Threads.stop) {
// Signal the cluster about stopping
stop_signal = true;
MPI_Ibarrier(StopComm, &reqStop);
MPI_Wait(&reqStop, MPI_STATUS_IGNORE);
}
}
else {
int flagStop;
// Check if we've received any stop signal
MPI_Test(&reqStop, &flagStop, MPI_STATUS_IGNORE);
if (flagStop)
Threads.stop = true;
}
}
int size() {
return world_size;
}