multicore support seems to work

diff --git a/main.c b/main.c
index a81687172b545275b1c5b10d1c27c8c14bc78873..38bd8541ca6d5c382aa6beb06cd5d088f18f6ed8 100644 (file)
--- a/main.c
+++ b/main.c
@@ -30,6 +30,10 @@
 #include <unistd.h>
 #include <stdbool.h>
 #include <inttypes.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <sys/uio.h>
+#include <sys/fcntl.h>
 
 #include <publib.h>
 #include <glpk.h>
@@ -89,13 +93,218 @@ static AdjWord *best_adjmatrix;
 
 static unsigned printcounter;
 
-static int ncpus = 1;
+static void iterate(void);
+static void iterate_recurse(int i, AdjWord min);
+static void optimise(bool doprint);
 
 static void progress_eol(void) {
   fprintf(stderr,"        \r");
   fflush(stderr);
 }
 
+/*----- multicore support -----*/
+
+/*
+ * Multicore protocol
+ *
+ * We fork into:
+ *   - master (parent)
+ *   - generator
+ *   - ncpu workers
+ *
+ * ipc facilities:
+ *   - one pipe ("work") from generator to workers
+ *   - ever-extending file ("bus") containing new "best" values
+ *   - one file for each worker giving maxhamweight and adjmatrix for best
+ *
+ * generator runs iterate_recurse to a certain depth and writes the
+ * candidates to a pipe
+ *
+ * workers read candidates from the pipe and resume iterate_recurse
+ * halfway through the recursion
+ *
+ * whenever a worker does a doprint, it checks the bus for new best
+ * value; actual best values are appended
+ *
+ * master waits for generator and all workers to finish and then
+ * runs optimise() for each worker's best, then prints
+ */ 
+
+static int ncpus = 0, multicore_iteration_boundary = INT_MAX;
+
+static int mc_bus, mc_work[2];
+static off_t mc_bus_read;
+
+typedef struct {
+  int w;
+  FILE *results;
+  pid_t pid;
+} Worker;
+static Worker *mc_us;
+
+#define MAX_NIOVS 3
+static AdjWord mc_iter_min;
+static int mc_niovs;
+static size_t mc_iovlen;
+static struct iovec mc_iov[MAX_NIOVS];
+
+#define IOV0 (mc_niovs = mc_iovlen = 0)
+
+#define IOV(obj, count) ({                             \
+    assert(mc_niovs < MAX_NIOVS);                      \
+    mc_iov[mc_niovs].iov_base = &(obj);                        \
+    mc_iov[mc_niovs].iov_len = sizeof(obj) * (count);  \
+    mc_iovlen += mc_iov[mc_niovs].iov_len;              \
+    mc_niovs++;                                                \
+  })
+
+static void mc_rwvsetup_outer(void) {
+  IOV0;
+  IOV(maxhamweight, 1);
+  IOV(mc_iter_min, 1);
+  IOV(*adjmatrix, multicore_iteration_boundary);
+}
+
+static void mc_rwvsetup_full(void) {
+  IOV0;
+  IOV(*adjmatrix, n);
+}
+
+static void vlprintf(const char *fmt, va_list al) {
+  vfprintf(stderr,fmt,al);
+  progress_eol();
+}
+
+static void LPRINTF(const char *fmt, ...) {
+  va_list al;
+  va_start(al,fmt);
+  vlprintf(fmt,al);
+  va_end(al);
+}
+
+static void mc_awaitpid(int wnum, pid_t pid) {
+  LPRINTF("master awaiting %2d [%ld]",wnum,(long)pid);
+  int status;
+  pid_t got = waitpid(pid, &status, 0);
+  assert(got == pid);
+  if (status) {
+    fprintf(stderr,"\nFAILED SUBPROC %2d [%ld] %d\n",
+           wnum, (long)pid, status);
+    exit(-1);
+  }
+}
+
+static void multicore_outer_iteration(int i, AdjWord min) {
+  assert(i == multicore_iteration_boundary);
+  mc_iter_min = min;
+  mc_rwvsetup_outer();
+  ssize_t r = writev(mc_work[1], mc_iov, mc_niovs);
+  assert(r == mc_iovlen);
+  /* effectively, this writev arranges to transfers control
+   * to some worker's instance of iterate_recurse via mc_iterate_worker */
+}
+
+static void mc_iterate_worker(void) {
+  for (;;) {
+    mc_rwvsetup_outer();
+    ssize_t r = readv(mc_work[0], mc_iov, mc_niovs);
+    if (r == 0) break;
+    assert(r == mc_iovlen);
+
+    /* stop iterate_recurse from trying to run multicore_outer_iteration */
+    int mc_org_it_bound = multicore_iteration_boundary;
+    multicore_iteration_boundary = INT_MAX;
+    iterate_recurse(mc_org_it_bound, mc_iter_min);
+    multicore_iteration_boundary = mc_org_it_bound;
+  }
+  LPRINTF("worker %2d reporting",mc_us->w);
+  if (best_adjmatrix) {
+    adjmatrix = best_adjmatrix;
+    mc_rwvsetup_full();
+    ssize_t r = writev(fileno(mc_us->results), mc_iov, mc_niovs);
+    assert(r == mc_iovlen);
+  }
+  LPRINTF("worker %2d ending",mc_us->w);
+  exit(0);
+}
+
+static void multicore(void) {
+  Worker *mc_workers;
+  int w;
+  pid_t genpid;
+
+  multicore_iteration_boundary = n / 2;
+
+  FILE *busf = tmpfile();  assert(busf);
+  mc_bus = fileno(busf);
+  int r = fcntl(mc_bus, F_GETFL);  assert(r >= 0);
+  r |= O_APPEND;
+  r = fcntl(mc_bus, F_SETFL, r);  assert(r >= 0);
+
+  r = pipe(mc_work);  assert(!r);
+
+  mc_workers = xmalloc(sizeof(*mc_workers) * ncpus);
+  for (w=0; w<ncpus; w++) {
+    mc_workers[w].w = w;
+    mc_workers[w].results = tmpfile();  assert(mc_workers[w].results);
+    mc_workers[w].pid = fork();  assert(mc_workers[w].pid >= 0);
+    if (!mc_workers[w].pid) {
+      mc_us = &mc_workers[w];
+      close(mc_work[1]);
+      LPRINTF("worker %2d running", w);
+      mc_iterate_worker();
+      exit(0);
+    }
+  }
+
+  close(mc_work[0]);
+
+  genpid = fork();  assert(genpid >= 0);
+  if (!genpid) {
+    LPRINTF("generator running");
+    iterate();
+    exit(0);
+  }
+
+  close(mc_work[1]);
+  mc_awaitpid(-1, genpid);
+  for (w=0; w<ncpus; w++)
+    mc_awaitpid(w, mc_workers[w].pid);
+
+  for (w=0; w<ncpus; w++) {
+    mc_rwvsetup_full();
+    LPRINTF("reading report from %2d",w);
+    ssize_t sr = preadv(fileno(mc_workers[w].results), mc_iov, mc_niovs, 0);
+    if (!sr) continue;
+    maxhamweight = 0;
+    optimise(1);
+  }
+}
+
+static void multicore_check_for_new_best(void) {
+  if (!ncpus) return;
+
+  for (;;) {
+    double msg;
+    ssize_t got = pread(mc_bus, &msg, sizeof(msg), mc_bus_read);
+    if (!got) break;
+    assert(got == sizeof(msg));
+    if (msg > best)
+      best = msg;
+    mc_bus_read += sizeof(msg);
+  }
+}
+
+static void multicore_found_new_best(void) {
+  if (!ncpus) return;
+
+  if (mc_us /* might be master */) fprintf(stderr,"    w%-2d ",mc_us->w);
+  ssize_t wrote = write(mc_bus, &best, sizeof(best));
+  assert(wrote == sizeof(best));
+}
+
+/*----- end of multicore support -----*/
+
 static AdjWord *xalloc_adjmatrix(void) {
   return xmalloc(sizeof(*adjmatrix)*n);
 }
@@ -323,6 +532,7 @@ static void optimise(bool doprint) {
   HAVE_PRINTED;
 
   best = got;
+  multicore_found_new_best();
 
   if (best_prob) glp_delete_prob(best_prob);
   best_prob = prob;
@@ -339,6 +549,7 @@ static void optimise(bool doprint) {
   if (prob)
     glp_delete_prob(prob);
   if (doprint) progress_eol();
+  if (doprint) multicore_check_for_new_best();
 }
 
 static void iterate_recurse(int i, AdjWord min) {
@@ -347,6 +558,10 @@ static void iterate_recurse(int i, AdjWord min) {
     optimise(!(printcounter & 0xfff));
     return;
   }
+  if (i >= multicore_iteration_boundary) {
+    multicore_outer_iteration(i, min);
+    return;
+  }
   for (adjmatrix[i] = min;
        ;
        adjmatrix[i]++) {
@@ -419,7 +634,10 @@ int main(int argc, char **argv) {
   m = atoi(argv[2]);
 
   prep();
-  iterate();
+
+  if (ncpus) multicore();
+  else iterate();
+
   report();
   return 0;
 }