Printer FriendlyEmail Article Link

Spirent TestCenter: What kind of latency can be expected with Cut-Through or Store-and-Forward type devices?

  • Spirent TestCenter
    • PGA
  • A switch can be a Cut-Through or Store-and-Forward type device.
    • With Cut-Through - the switch starts forwarding traffic when it receives the first 6 bytes, the destination MAC address, and knows where the traffic needs to go to.
      • With Cut-Through devices, the frame size doesn't affect the lattency measurements
    • With Store-and-Forward - the switch buffers the whole frame and then forwards it on to the destination.
      • Store-and-forward in most cases is slower, for example it takes ~1200 nanoseconds to put a 1500 byte frame on the wire at 10G speeds.
      • As per the frame size increases (64...512...1518... and so on) you would expect more latency. Higher frame size = Higher lattency
      • Spirent TestCenter RFC tests can remove the 1200 nanoseconds to only measure the switch time by putting it in LIFO (store-and-forward) mode.
    • Going from 40G to 10G there is a speed mismatch (not media conversion) which in most cases means the switch has to store-and-forward, ie. can’t cut-through 40Gbps to 10Gbps, you’d lose 30Gbps…
  • It is important to know what latency measurement mode is used.
    • When comparing latencies it is best to always use FIFO on Spirent TestCenter for both Cut-Through and Store-and-Forward configurations.
    • The RFC 2544 latency test is a great way to easily run these tests.
  • If the latency shows to always be 800ns going from 40G to 40G and 650ns going from 40G to 10G indicates Spirent TestCenter was configured in LIFO store-and-forward mode.
    • As mentioned in this mode it correctly (per RFC definitions) subtracts the time it takes to put the frame on the wire and buffer it in the switch. This could be why the number is lower going from 40G to 10G.
  • You can also get the back to back latency to almost 0 by removing latency compensation in Spirent TestCenter:

Following papers are great reads with lots of detail on testing this type of switch:

  • Here is some additional information on that topic:
    • It is possible to set the Latency Timestamp Reference Location different of each side. However, for most test cases, both sides should be set to Start of Frame which is the default.
    • So really, why are there different Latency Timestamp Reference Locations?
      • In summary; so we can measure/display Store-and-Forward latency in real time, separately for potentially 64,000 individual streams, and at up to 150 million frames per second!
        To do this you would set Generator side to End-of-Frame and the Analyzer side to Start-of-Frame
        This essentially would subtract out the frame bit time as per RFC 2544 for measuring Store-and-Forward latency
        However, prior to Spirent TestCenter this was always done posttest in software
        Now it can been done by the hardware, and displayed, in real time!


Product : Spirent TestCenter