EVALUATION

http://hg.openjdk.java.net/jdk7/hotspot/hotspot/rev/3698e8f47799
                                     

EVALUATION

http://hg.openjdk.java.net/jdk7/hotspot-gc/hotspot/rev/3698e8f47799
                                     

SUGGESTED FIX

While investigating this failure I also noticed the following:

  update_recent_gc_times(end_sec, full_gc_time_sec);

This is incorrect as update_recent_gc_time() is expecting the second parameter to be in ms:

void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
                                               double elapsed_ms) {

The fix is straightforward (as full_gc_time_ms is already calculated):

-  update_recent_gc_times(end_sec, full_gc_time_sec);
+  update_recent_gc_times(end_sec, full_gc_time_ms);
                                     

SUGGESTED FIX

Revert this change:

+     if (!evacuation_failed()) {
        g1_policy()->record_pause_time((end_time_sec - start_time_sec)*1000.0);
+     }

to make sure we correctly set _last_pause_ms, but only update the pause statistics at the end of the pause if the pause didn't have evacuation failure (which was the original intention of the change).
                                     

EVALUATION

Even though it looks similar to 6785211 I think the cause is different. The failure in this CR was introduced by 6797754 and in particular this change:

+     if (!evacuation_failed()) {
        g1_policy()->record_pause_time((end_time_sec - start_time_sec)*1000.0);
+     }

So, if there was an evacuation failure, record_pause_time() was not called to notify the policy how long the pause time was. As a result the policy will pick up junk if it reads _last_pause_ms (which is the case here), or pick up an incorrect value (if _last_pause_ms) had been set by a previous pause.