ATM bandwidth

Unanswered Question
Jun 3rd, 2008
User Badges:

Issue: we recently installed an ATM DS3 circuit. We contracted with our telecom provider for 15 Mbits. I configured the vbr-nrt at 15000 (per the telecom engineer) on each end of the pvc circuit. However, when I test the circuit throughput with iperf I am only getting 12.2 MBits. The telecom provider says this is in line with expectations. I know ATM will have additional overhead than other protocols, but 19% seems like a lot to me. Can anyone provide additional insight into this?


Thanks in advance.

  • 1
  • 2
  • 3
  • 4
  • 5
Overall Rating: 4.3 (3 ratings)
Loading.
paolo bevilacqua Tue, 06/03/2008 - 10:01
User Badges:
  • Super Gold, 25000 points or more
  • Hall of Fame,

    Founding Member

Yes, these exactly are the figures. See:

http://sd.wareonearth.com/~phil/net/overhead/


Also, when measuring with tcperf, performances are dependent by the receive window size used. But in your case, cell tax is what is eating you the bandwidth.

Giuseppe Larosa Tue, 06/03/2008 - 10:46
User Badges:
  • Super Silver, 17500 points or more
  • Hall of Fame,

    Founding Member

Hello Paolo,

very nice link I've added it to my bookmarks.


Also the segmentation and reassembly of packets into an AALx PDU and the into cells and back into an AALx PDU takes its time.


I would suggest to try to use two concurrent sessions of tcperf to see if there are any changes but probably nothing meaningful in this case.

Some years ago for a 100 Mbps test over two parallel STM-1 ATM links I remember that using two concurrent tcperf sessions led to better results and we thought it was for SAR activity.


Best Regards

Giuseppe

Joseph W. Doherty Tue, 06/03/2008 - 16:41
User Badges:
  • Super Bronze, 10000 points or more

Two items to be aware of with ATM that consume bandidth are ATM cell headers, 5 bytes of 53 or 9.4% overhead, and the fact that cells are fixed sized, always 53 bytes. For instance if you had a 49 bytes, it would use two cells, 106 bytes so we get 53.8% overhead. So, best overhead is "only" 9.4%, worst overhead is 53.8%, and perhaps an average overhead of about 15 to 20%.

Actions

This Discussion