Use server receive rate rather than client send()/buffering rate

Fixes #412

mininet/net.py

@@ -689,7 +689,10 @@ class Mininet( object ):
         """Run iperf between two hosts.
            hosts: list of hosts; if None, uses opposite hosts
            l4Type: string, one of [ TCP, UDP ]
-           returns: results two-element array of server and client speeds"""
+           returns: results two-element array of [ server, client ] speeds
+           note: send() is buffered, so client rate can be much higher than
+           the actual transmission rate; on an unloaded system, server
+           rate should be much closer to the actual receive rate"""
         if not quietRun( 'which telnet' ):
             error( 'Cannot find telnet in $PATH - required for iperf test' )
             return

mininet/test/test_hifi.py

@@ -136,9 +136,14 @@ class testOptionsTopoCommon( object ):
                 'switch = %s\n'
                 % ( BW, bw_strs, N, loptsStr, self.switchClass ) )
 
-        for bw_str in bw_strs:
-            bw = float( bw_str.split(' ')[0] )
-            self.assertWithinTolerance( bw, BW, BW_TOLERANCE, msg )
+        # On the client side, iperf doesn't wait for ACKs - it simply
+        # reports how long it took to fill up the TCP send buffer.
+        # As long as the kernel doesn't wait a long time before
+        # delivering bytes to the iperf server, its reported data rate
+        # should be close to the actual receive rate.
+        serverRate, clientRate = bw_strs
+        bw = float( serverRate.split(' ')[0] )
+        self.assertWithinTolerance( bw, BW, BW_TOLERANCE, msg )
 
     def testLinkDelay( self ):
         "Verify that link delays are accurate within a bound."