sort: do not generate thousands of subprocesses for 16-way merge

Without this change, tests/misc/sort-compress-hang would consume
more than 10,000 process slots on my RHEL 5.5 x86-64 server,
making it likely for other applications to fail due to lack of
process slots.  With this change, the same benchmark causes 'sort'
to consume at most 19 process slots.  The change also improved
wall-clock time by 2% and user+system time by 14% on that benchmark.
* NEWS: Document this.
* src/sort.c (MAX_PROCS_BEFORE_REAP): Remove.
(reap_exited): Renamed from reap_some; this is a more accurate name,
since "some" incorrectly implies that it reaps at least one process.
All uses changed.
(reap_some): New function: it *does* reap at least one process.
(pipe_fork): Do not allow more than NMERGE + 2 subprocesses.
(mergefps, sort): Omit check for exited processes: no longer needed,
and anyway the code consumed too much CPU per line when 2 < nprocs.

1 files changed, 21 insertions, 13 deletions

diff --git a/src/sort.c b/src/sort.c
index 6bce49b4e..54dd81528 100644
--- a/src/sort.c
+++ b/src/sort.c
@@ -659,9 +659,6 @@ proctab_comparator (void const *e1, void const *e2)
 /* The number of unreaped child processes.  */
 static pid_t nprocs;
 
-/* The number of child processes we'll allow before we try to reap some. */
-enum { MAX_PROCS_BEFORE_REAP = 2 };
-
 static bool delete_proc (pid_t);
 
 /* If PID is positive, wait for the child process with that PID to
@@ -743,12 +740,21 @@ wait_proc (pid_t pid)
    already exited.  */
 
 static void
-reap_some (void)
+reap_exited (void)
 {
   while (0 < nprocs && reap (0))
     continue;
 }
 
+/* Reap at least one exited child, waiting if necessary.  */
+
+static void
+reap_some (void)
+{
+  reap (-1);
+  reap_exited ();
+}
+
 /* Reap all children, waiting if necessary.  */
 
 static void
@@ -969,6 +975,16 @@ pipe_fork (int pipefds[2], size_t tries)
   if (pipe (pipefds) < 0)
     return -1;
 
+  /* At least NMERGE + 1 subprocesses are needed.  More could be created, but
+     uncontrolled subprocess generation can hurt performance significantly.
+     Allow at most NMERGE + 2 subprocesses, on the theory that there
+     may be some useful parallelism by letting compression for the
+     previous merge finish (1 subprocess) in parallel with the current
+     merge (NMERGE + 1 subprocesses).  */
+
+  if (nmerge + 1 < nprocs)
+    reap_some ();
+
   while (tries--)
     {
       /* This is so the child process won't delete our temp files
@@ -991,7 +1007,7 @@ pipe_fork (int pipefds[2], size_t tries)
         {
           xnanosleep (wait_retry);
           wait_retry *= 2;
-          reap_some ();
+          reap_exited ();
         }
     }
 
@@ -2968,10 +2984,6 @@ mergefps (struct sortfile *files, size_t ntemps, size_t nfiles,
           ord[j] = ord[j + 1];
         ord[count_of_smaller_lines] = ord0;
       }
-
-      /* Free up some resources every once in a while.  */
-      if (MAX_PROCS_BEFORE_REAP < nprocs)
-        reap_some ();
     }
 
   if (unique && savedline)
@@ -3829,10 +3841,6 @@ sort (char *const *files, size_t nfiles, char const *output_file,
 
           xfclose (tfp, temp_output);
 
-          /* Free up some resources every once in a while.  */
-          if (MAX_PROCS_BEFORE_REAP < nprocs)
-            reap_some ();
-
           if (output_file_created)
             goto finish;
         }