Skip to Content.
Sympa Menu

ndt-users - Re: Overoptimistic estimates

Subject: ndt-users list created

List archive

Re: Overoptimistic estimates


Chronological Thread 
  • From: Richard Carlson <>
  • To: Jon Kåre Hellan <>,
  • Subject: Re: Overoptimistic estimates
  • Date: Fri, 17 Feb 2006 11:21:44 -0500

Hi Jon;

Yes you are correct. The NDT server uses the libpcap timestamp to calculate the bottleneck link speed, and yes interrupt coalescing will skew these results. The problem is the timestamps are added to the packet when the CPU see's them. With coalescing on, multiple packets arrive at PCI bus or memory I/O speeds and this skews the results.

You should try turning off the coalescing function, while this will raise the CPU load it should not be a major problem for a dedicated NDT server. The e1000.txt file describes how to set the RxIntDelay and RxAbsIntDelay values to control the coalescing function.

I will update the NDT documentation to point out this hazard.

Of course the ideal solution would be to get the NIC to timestamp the arriving packets, but I've only heard of the Syskonnect cards actually doing this.

Thanks for the feedback.
Rich

At 04:36 AM 2/17/2006, Jon Kåre Hellan wrote:
Jon Kåre Hellan wrote:
Hi,
I've placed ndt servers at various points in our network, but find that way too many links are reported as 10 Gbps.
Here's part of a web100srv logfile:
Web100srv (ver 3.1.4) process (9461) started Feb 13 14:36:51
Feb 13 14:36:53 persaunet.uninett.no port 49412
spds[0] = ' 0 0 300 159 2105 5281 53 6 30 7632 0 0 38.68' max=7632 [49.03%]

I've looked at the source and found out how NDT estimates link speeds. It uses packet pairs, and places speed observations in bins. The bin with the most observations wins. In this case it's bin 9, with count 7632. It represents the range 2500-10000 Mbps.

The interface cards are Intel Gigabit Ethernet - 82547GI according to lspci. Could interrupt coalescing be the cause of the strange results? May be we are actually measuring bus speed.

Regards

Jon

------------------------------------



Richard A. Carlson e-mail:

Network Engineer phone: (734) 352-7043
Internet2 fax: (734) 913-4255
1000 Oakbrook Dr; Suite 300
Ann Arbor, MI 48104



Archive powered by MHonArc 2.6.16.

Top of Page