changed the double send buffers to a chained buffer using iovec send operations. buffers are recycled (saves memory and cpu) and memory copies are avoided (saves cpu). More statistics is added to track buffer usage

parse_buffer_log.py

@@ -0,0 +1,83 @@
+import os, sys, time
+
+lines = open(sys.argv[1], 'rb').readlines()
+
+#keys = ['send_buffer_utilization']
+keys = ['send_buffer_size', 'used_send_buffer', 'protocol_buffer']
+#keys = ['send_buffer_alloc', 'send_buffer', 'allocate_buffer_alloc', 'allocate_buffer', 'protocol_buffer']
+#keys = ['send_buffer_alloc', 'send_buffer', 'allocate_buffer_alloc', 'allocate_buffer', 'protocol_buffer', 'append_send_buffer']
+
+average = ['send_buffer_utilization', 'send_buffer_size', 'used_send_buffer']
+average_interval = 120000
+render = 'lines'
+
+time_limit = -1
+if len(sys.argv) > 2:
+   time_limit = long(sys.argv[2])
+
+
+# logfile format:
+# <time(ms)> <key> <value>
+# example:
+# 16434 allocate_buffer: 17
+for k in keys:
+
+   last_sample = 0
+   average_accumulator = 0
+   average_samples = 0
+   peak = 0
+
+   out = open(k + '.dat', 'wb')
+   eval_average = False
+   if k in average:
+      eval_average = True
+      peak_out = open(k + '_peak.dat', 'wb')
+
+   for l in lines:
+      l = l.split(' ')
+      if len(l) != 3:
+         print l
+         continue
+      try:
+         if l[1] == k + ':':
+            if time_limit != -1 and long(l[0]) > time_limit: break
+            time = l[0]
+            value = l[2]
+            if eval_average:
+               while long(time) > last_sample + average_interval:
+                  last_sample = last_sample + average_interval
+                  if average_samples < 1: average_samples = 1
+                  print >>out, '%d %f' % (last_sample, average_accumulator / average_samples)
+                  print >>peak_out, '%d %f' % (last_sample, peak)
+                  average_accumulator = 0
+                  average_samples = 0
+                  peak = 0
+               average_accumulator = average_accumulator + float(value)
+               average_samples = average_samples + 1
+               if float(value) > peak: peak = float(value)
+            else:
+               print >>out, time + ' ' + value,
+      except:
+         print l
+
+
+out = open('send_buffer.gnuplot', 'wb')
+print >>out, "set term png size 1200,700"
+print >>out, 'set output "send_buffer.png"'
+print >>out, 'set xrange [0:*]'
+print >>out, 'set xlabel "time (ms)"'
+print >>out, 'set ylabel "bytes (B)"'
+print >>out, "set style data lines"
+print >>out, "set key box"
+print >>out, 'plot',
+for k in keys:
+   if k in average:
+      print >>out, ' "%s.dat" using 1:2 title "%s %d seconds average" with %s,' % (k, k, average_interval / 1000., render),
+      print >>out, ' "%s_peak.dat" using 1:2 title "%s %d seconds peak" with %s,' % (k, k, average_interval / 1000., render),
+   else:
+      print >>out, ' "%s.dat" using 1:2 title "%s" with %s,' % (k, k, render),
+print >>out, 'x=0'
+out.close()
+
+os.system('gnuplot send_buffer.gnuplot');
+