Tune BWE to be a bit less sensitive to spurious delay events.

webrtc/modules/remote_bitrate_estimator/overuse_detector.cc

Index: webrtc/modules/remote_bitrate_estimator/overuse_detector.cc
diff --git a/webrtc/modules/remote_bitrate_estimator/overuse_detector.cc b/webrtc/modules/remote_bitrate_estimator/overuse_detector.cc
index 042fe9aa2ae05625ce88861225637fc1bd1278aa..2f1ce36c303e9aacda0535a873a2376d3e58b6d2 100644
--- a/webrtc/modules/remote_bitrate_estimator/overuse_detector.cc
+++ b/webrtc/modules/remote_bitrate_estimator/overuse_detector.cc
@@ -142,8 +142,10 @@ void OveruseDetector::UpdateThreshold(double modified_offset, int64_t now_ms) {
   }
 
   const double k = fabs(modified_offset) < threshold_ ? k_down_ : k_up_;
+  const int64_t kMaxTimeDeltaMs = 100;

[stefan-webrtc, 2016/03/10 11:34:41]

This was caught by the existing tests:
TestShortTimeoutAndWrap
TestLongTimeoutAndWrap

They didn't catch it previously because they ran for too long after the timeout
so that the (quicker) filter caught up. Reduced the run time after the timeout
so that we are sure to catch regressions.

[mflodman, 2016/03/10 11:59:33]

Good catch!

[mflodman, 2016/03/10 11:59:33]

After discussions, I agree that 100 is a good threshold to start with and we can
experiment with higher values later. This will rather err on the pre-fairness
behavior and would be a safer step to take.

+  int64_t time_delta_ms = std::min(now_ms - last_update_ms_, kMaxTimeDeltaMs);
   threshold_ +=
-      k * (fabs(modified_offset) - threshold_) * (now_ms - last_update_ms_);
+      k * (fabs(modified_offset) - threshold_) * time_delta_ms;
 
   const double kMinThreshold = 6;
   const double kMaxThreshold = 600;