Throughput Testing

How fast can the network go? What is the actual usable capacity of a particular network link? You can get a very good estimate of your throughput capacity by flooding the link with traffic and measuring how long it takes to transfer the data. While there are web pages available that will perform a “speed test” in your browser (such as www.dslreports.com/stest), these tests are increasingly inaccurate as you get further from the testing source. Even worse, they do not allow you to test the speed of a particular link, but only the speed of your link to the Internet. Here are two tools that will allow you to perform throughput testing on your own networks.

• ttcp (ftp.arl.mil/ftp/pub/ttcp/). Now a standard part of most Unix-like systems, ttcp is a simple network performance testing tool. One instance is run on either side of the link you want to test. The first node runs in receive mode, and the other transmits:

  node_a$ ttcp -r -s
  
  node_b$ ttcp -t -s node_attcp-t: buflen=8192, nbuf=2048, align=16384/0,
                                   port=5001 tcp -> node_attcp-t: socket
  
  ttcp-t: connect
  ttcp-t: 16777216 bytes in 249.14 real seconds = 65.76 KB/sec +++
  ttcp-t: 2048 I/O calls, msec/call = 124.57, calls/sec = 8.22
  ttcp-t: 0.0user 0.2sys 4:09real 0% 0i+0d 0maxrss 0+0pf 7533+0csw
  

  After collecting data in one direction, you should reverse the transmit and receive partners to test the link in the other direction. It can test UDP as well as TCP streams, and can alter various TCP parameters and buffer lengths to give the network a good workout. It can even use a user-supplied data stream instead of sending random data. Remember that the speed readout is in kilobytes, not kilobits. Multiply the result by 8 to find the speed in kilobits per second.

  The only real disadvantage to ttcp is that it hasn't been developed in years. Fortunately, the code has been released in the public domain and is freely available. Like ping and traceroute, ttcp is found as a standard tool on many systems.

• iperf (dast.nlanr.net/Projects/Iperf/). Much like ttcp, iperf is a commandline tool for estimating the throughput of a network connection. It supports many of the same features as ttcp, but uses a “client” and “server” model instead of a “receive” and “transmit” pair. To run iperf, launch a server on one side and a client on the other:

  node_a$ iperf -s
  
  node_b$ iperf -c node_a
  
  Client connecting to node_a, TCP port 5001
  TCP window size: 16.0 KByte (default)
  [  5] local 10.15.6.1 port 1212 connected with 10.15.6.23 port 5001
  [ ID] Interval Transfer Bandwidth
  [  5] 0.0-11.3 sec 768 KBytes 558 Kbits/sec
  

  The server side will continue to listen and accept client connections on port 5001 until you hit control-C to kill it. This can make it handy when running multiple test runs from a variety of locations.

  The biggest difference between ttcp and iperf is that iperf is under active development, and has many new features (including IPv6 support). This makes it a good choice as a performance tool when building new networks.