streamthinca: add latency_tolerance feature

- allow coinc processing to be deferred until the oldest trigger in the
  queues reaches some minimum age to condense coinc processing into fewer,
  less frequent, intervals at the expense of latency

gstlal-inspiral/python/streamthinca.py

@@ -235,7 +235,7 @@ class timereversebackgroundcollector(object):
 
 
 class StreamThinca(object):
-	def __init__(self, xmldoc, process_id, delta_t, min_instruments = 2, sngls_snr_threshold = None, background_collector_type = "normal"):
+	def __init__(self, xmldoc, process_id, delta_t, min_instruments = 2, sngls_snr_threshold = None, background_collector_type = "normal", latency_tolerance = 0.):
 		if background_collector_type not in ("normal", "time_reverse"):
 			raise ValueError("background_collector_type can only be 'normal' or 'time_reverse'")
 		self.background_collector_type = background_collector_type
@@ -243,6 +243,7 @@ class StreamThinca(object):
 		self.delta_t = delta_t
 		self.min_instruments = min_instruments
 		self.sngls_snr_threshold = sngls_snr_threshold
+		self.latency_tolerance = latency_tolerance
 		self.set_xmldoc(xmldoc, process_id)
 		self.clustered_sngl_ids = set()
 
@@ -267,9 +268,17 @@ class StreamThinca(object):
 		Push new triggers into the coinc engine.  Returns True if
 		the coinc engine's internal state has changed in a way that
 		might enable new candidates to be constructed, False if
-		not.
+		not.  If latency_tolerance is not 0, then triggers
+		are allowed to accumulate in the internal queues until the
+		oldest is latency_tolerance older than t_complete before
+		this method reports that candidates can be extracted from
+		the graph.  This concentrates the time spent in the Python
+		coincidence code into smaller intervals, which allows the
+		gstreamer code to achieve greater parallelism, which
+		achieves more efficient CPU use at the expense of longer
+		latencies.
 		"""
-		return self.time_slide_graph.push(instrument, events, t_complete)
+		return self.time_slide_graph.push(instrument, events, t_complete) and (t_complete - self.time_slide_graph.age() >= self.latency_tolerance)
 
 
 	def pull(self, rankingstat, fapfar = None, zerolag_rankingstatpdf = None, coinc_sieve = None, flush = False, cluster = False, cap_singles = False, FAR_trialsfactor = 1.0, template_id_time_map = None):