6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads

Select number of GC threads dynamically based on heap usage and number of Java threads

Reviewed-by: johnc, ysr, jcoomes

hotspot/src/share/vm/gc_implementation/shared/adaptiveSizePolicy.cpp

@@ -28,8 +28,10 @@
 #include "memory/collectorPolicy.hpp"
 #include "runtime/timer.hpp"
 #include "utilities/ostream.hpp"
+#include "utilities/workgroup.hpp"
 elapsedTimer AdaptiveSizePolicy::_minor_timer;
 elapsedTimer AdaptiveSizePolicy::_major_timer;
+bool AdaptiveSizePolicy::_debug_perturbation = false;
 
 // The throughput goal is implemented as
 //      _throughput_goal = 1 - ( 1 / (1 + gc_cost_ratio))
@@ -88,6 +90,134 @@ AdaptiveSizePolicy::AdaptiveSizePolicy(size_t init_eden_size,
   _young_gen_policy_is_ready = false;
 }
 
+//  If the number of GC threads was set on the command line,
+// use it.
+//  Else
+//    Calculate the number of GC threads based on the number of Java threads.
+//    Calculate the number of GC threads based on the size of the heap.
+//    Use the larger.
+
+int AdaptiveSizePolicy::calc_default_active_workers(uintx total_workers,
+                                            const uintx min_workers,
+                                            uintx active_workers,
+                                            uintx application_workers) {
+  // If the user has specifically set the number of
+  // GC threads, use them.
+
+  // If the user has turned off using a dynamic number of GC threads
+  // or the users has requested a specific number, set the active
+  // number of workers to all the workers.
+
+  uintx new_active_workers = total_workers;
+  uintx prev_active_workers = active_workers;
+  uintx active_workers_by_JT = 0;
+  uintx active_workers_by_heap_size = 0;
+
+  // Always use at least min_workers but use up to
+  // GCThreadsPerJavaThreads * application threads.
+  active_workers_by_JT =
+    MAX2((uintx) GCWorkersPerJavaThread * application_workers,
+         min_workers);
+
+  // Choose a number of GC threads based on the current size
+  // of the heap.  This may be complicated because the size of
+  // the heap depends on factors such as the thoughput goal.
+  // Still a large heap should be collected by more GC threads.
+  active_workers_by_heap_size =
+      MAX2((size_t) 2U, Universe::heap()->capacity() / HeapSizePerGCThread);
+
+  uintx max_active_workers =
+    MAX2(active_workers_by_JT, active_workers_by_heap_size);
+
+  // Limit the number of workers to the the number created,
+  // (workers()).
+  new_active_workers = MIN2(max_active_workers,
+                                (uintx) total_workers);
+
+  // Increase GC workers instantly but decrease them more
+  // slowly.
+  if (new_active_workers < prev_active_workers) {
+    new_active_workers =
+      MAX2(min_workers, (prev_active_workers + new_active_workers) / 2);
+  }
+
+  // Check once more that the number of workers is within the limits.
+  assert(min_workers <= total_workers, "Minimum workers not consistent with total workers");
+  assert(new_active_workers >= min_workers, "Minimum workers not observed");
+  assert(new_active_workers <= total_workers, "Total workers not observed");
+
+  if (ForceDynamicNumberOfGCThreads) {
+    // Assume this is debugging and jiggle the number of GC threads.
+    if (new_active_workers == prev_active_workers) {
+      if (new_active_workers < total_workers) {
+        new_active_workers++;
+      } else if (new_active_workers > min_workers) {
+        new_active_workers--;
+      }
+    }
+    if (new_active_workers == total_workers) {
+      if (_debug_perturbation) {
+        new_active_workers =  min_workers;
+      }
+      _debug_perturbation = !_debug_perturbation;
+    }
+    assert((new_active_workers <= (uintx) ParallelGCThreads) &&
+           (new_active_workers >= min_workers),
+      "Jiggled active workers too much");
+  }
+
+  if (TraceDynamicGCThreads) {
+     gclog_or_tty->print_cr("GCTaskManager::calc_default_active_workers() : "
+       "active_workers(): %d  new_acitve_workers: %d  "
+       "prev_active_workers: %d\n"
+       " active_workers_by_JT: %d  active_workers_by_heap_size: %d",
+       active_workers, new_active_workers, prev_active_workers,
+       active_workers_by_JT, active_workers_by_heap_size);
+  }
+  assert(new_active_workers > 0, "Always need at least 1");
+  return new_active_workers;
+}
+
+int AdaptiveSizePolicy::calc_active_workers(uintx total_workers,
+                                            uintx active_workers,
+                                            uintx application_workers) {
+  // If the user has specifically set the number of
+  // GC threads, use them.
+
+  // If the user has turned off using a dynamic number of GC threads
+  // or the users has requested a specific number, set the active
+  // number of workers to all the workers.
+
+  int new_active_workers;
+  if (!UseDynamicNumberOfGCThreads ||
+     (!FLAG_IS_DEFAULT(ParallelGCThreads) && !ForceDynamicNumberOfGCThreads)) {
+    new_active_workers = total_workers;
+  } else {
+    new_active_workers = calc_default_active_workers(total_workers,
+                                                     2, /* Minimum number of workers */
+                                                     active_workers,
+                                                     application_workers);
+  }
+  assert(new_active_workers > 0, "Always need at least 1");
+  return new_active_workers;
+}
+
+int AdaptiveSizePolicy::calc_active_conc_workers(uintx total_workers,
+                                                 uintx active_workers,
+                                                 uintx application_workers) {
+  if (!UseDynamicNumberOfGCThreads ||
+     (!FLAG_IS_DEFAULT(ConcGCThreads) && !ForceDynamicNumberOfGCThreads)) {
+    return ConcGCThreads;
+  } else {
+    int no_of_gc_threads = calc_default_active_workers(
+                             total_workers,
+                             1, /* Minimum number of workers */
+                             active_workers,
+                             application_workers);
+    return no_of_gc_threads;
+  }
+}
+
 bool AdaptiveSizePolicy::tenuring_threshold_change() const {
   return decrement_tenuring_threshold_for_gc_cost() ||
          increment_tenuring_threshold_for_gc_cost() ||

hotspot/src/share/vm/gc_implementation/shared/adaptiveSizePolicy.hpp

@@ -187,6 +187,8 @@ class AdaptiveSizePolicy : public CHeapObj {
   julong _young_gen_change_for_minor_throughput;
   julong _old_gen_change_for_major_throughput;
 
+  static const uint GCWorkersPerJavaThread  = 2;
+
   // Accessors
 
   double gc_pause_goal_sec() const { return _gc_pause_goal_sec; }
@@ -331,6 +333,8 @@ class AdaptiveSizePolicy : public CHeapObj {
   // Return true if the policy suggested a change.
   bool tenuring_threshold_change() const;
 
+  static bool _debug_perturbation;
+
  public:
   AdaptiveSizePolicy(size_t init_eden_size,
                      size_t init_promo_size,
@@ -338,6 +342,31 @@ class AdaptiveSizePolicy : public CHeapObj {
                      double gc_pause_goal_sec,
                      uint gc_cost_ratio);
 
+  // Return number default  GC threads to use in the next GC.
+  static int calc_default_active_workers(uintx total_workers,
+                                         const uintx min_workers,
+                                         uintx active_workers,
+                                         uintx application_workers);
+
+  // Return number of GC threads to use in the next GC.
+  // This is called sparingly so as not to change the
+  // number of GC workers gratuitously.
+  //   For ParNew collections
+  //   For PS scavenge and ParOld collections
+  //   For G1 evacuation pauses (subject to update)
+  // Other collection phases inherit the number of
+  // GC workers from the calls above.  For example,
+  // a CMS parallel remark uses the same number of GC
+  // workers as the most recent ParNew collection.
+  static int calc_active_workers(uintx total_workers,
+                                 uintx active_workers,
+                                 uintx application_workers);
+
+  // Return number of GC threads to use in the next concurrent GC phase.
+  static int calc_active_conc_workers(uintx total_workers,
+                                      uintx active_workers,
+                                      uintx application_workers);
+
   bool is_gc_cms_adaptive_size_policy() {
     return kind() == _gc_cms_adaptive_size_policy;
   }