Start in on the MPI driver

Makefile

@@ -64,7 +64,7 @@ dpnode: dpnode.o evutils.o dpnode_main.o
 worker: worker.o evutils.o cudadl.o worker_main.o
 	g++ -g -Wall $^ -o $@ -L$(LIBEVENT)/lib -Wl,-rpath=$(LIBEVENT)/lib -levent -levent_pthreads -lntl -L$(GMP) -lgmp -lpthread -L$(CUDA)/lib64 -lcudart
 
-mpi: mpi.o
+mpi: mpi.o controller.o evutils.o
 	mpiCC -g -Wall $^ -o $@ -L$(LIBEVENT)/lib $(LIBEVENT)/lib/libevent.a $(LIBEVENT)/lib/libevent_pthreads.a -lntl -L$(GMP) -lgmp -lpthread -L$(CUDA)/lib64 -lcudart
 
 mpi.o: mpi.cc

mpi.cc

@@ -1,6 +1,51 @@
 #include <stdio.h>
 #include <mpi.h>
 
+#include <NTL/ZZ.h>
+
+#include "controller.h"
+
+NTL_CLIENT
+
+static int mpi_size;
+
+static int controllerfds[2];
+
+static void boundcb(const char *boundaddr, unsigned short boundport)
+{
+    // Write the port and addr to the pipe
+    write(controllerfds[1], &boundport, 2);
+    write(controllerfds[1], boundaddr, strlen(boundaddr));
+}
+
+void desired_resources(const ZZ &order, unsigned short &desired_dpnodes,
+    unsigned int &max_workers, unsigned int &dpfreq)
+{
+    // How many DPnodes should we use for a problem of this size?
+    desired_dpnodes = 2;
+    // How many workers would we like to use?
+    ZZ sorder = SqrRoot(order >> 46);
+    if (NumBits(sorder) > 30) {
+	// Just use all the workers we can find
+	max_workers = 4294967295U;  // 2^32 - 1
+    } else {
+	max_workers = trunc_long(sorder,31) + 1;
+    }
+
+    // By default, 1 in 1000 points are distinguihed points.  The
+    // number in the next line is 2^32/1000
+    dpfreq = 4294967;
+    if (order < 1000) {
+	// Just make every point a DP
+	dpfreq = 4294967295U;
+    } else if (NumBits(order) < 27) {
+	// The frequency of DPs should be 10/sqrt(order) to avoid
+	// a DP-free cycle, so dpfreq = (10*2^32)/sqrt(order)
+	ZZ f = (to_ZZ(10) << 32) / SqrRoot(order);
+	dpfreq = trunc_long(f, 31);
+    }
+}
+
 int main(int argc, char **argv)
 {
     // Init MPI
@@ -9,12 +54,42 @@ int main(int argc, char **argv)
     char hostname[257];
     gethostname(hostname, 256);
 
-    int rank, size;
+    int rank;
 
     MPI_Comm_rank(MPI_COMM_WORLD, &rank);
-    MPI_Comm_size(MPI_COMM_WORLD, &size);
+    MPI_Comm_size(MPI_COMM_WORLD, &mpi_size);
+
+    if (rank == 0) {
+	// Start the controller
+	pipe(controllerfds);
+	if (fork() == 0) {
+	    // Child; close the read half of the pipe
+	    close(controllerfds[0]);
+	    unsigned short bindport;
+	    Worklist worklist;
+
+	    if (controller_parse_args(argc, argv, bindport, worklist)) {
+		std::cerr << "Usage: " << argv[0] << " [-p listenport] N1 iter1 N2 iter2 ...\n";
+		return 1;
+	    }
+
+	    return controller_main(worklist, bindport, boundcb);
+	} else {
+	    // Parent; close the write half of the pipe
+	    close(controllerfds[1]);
+	    unsigned short boundport;
+	    unsigned char boundaddr[257];
+	    int res;
+
+	    res = read(controllerfds[0], &boundport, 2);
+	    if (res < 2) return 1;
+	    res = read(controllerfds[0], boundaddr, 256);
+	    if (res < 1) return 1;
+	    boundaddr[res] = '\0';
 
-    printf("%s %d %d %d\n", hostname, getpid(), rank, size);
+	    std::cerr << "Child bound to " << boundaddr << ":" << boundport << "\n";
+	}
+    }
 
     MPI_Finalize();