Skip to Content.
Sympa Menu

ndt-users - Re: NDT reports packet reorrdering

Subject: ndt-users list created

List archive

Re: NDT reports packet reorrdering


Chronological Thread 
  • From: Richard Carlson <>
  • To: Matt Crawford <>,
  • Subject: Re: NDT reports packet reorrdering
  • Date: Thu, 16 Mar 2006 13:42:09 -0500

Matt;

I see a lot of NDT reports non-zero out-of-order stats. Much more than I ever expected. I've pretty well convinced myself that this is a host stack/driver issue instead of a network switch/router issue. My working hypothesis is that the NIC driver caused this by somehow re-ordering the packets before they were handed up to the kernel.

As far as the NDT server goes, out-of-order = DupAcksIn/AckPktsIn So the question is, why would the receiver generate DupAcks? As far as I know, the only reason is because a gap appeared in the TCP stream.

In any case, a small amount of reordering, if it doesn't lead to timeout or fast retransmits should not cause that big of an impact on a TCP flow. The receiver will just send a few more ACK packets. The web100 variable list should tell you if this is happening.

How low is this throughput? What does the BW*D product say using the send/receive buffers? What OS are you using. (I've had a bunch of problems with the Linux 2.6.13 and above NIC drivers. The e100.c file was causing me a tremendous amount of problems (tcptrace reported an initial TCP window of 42 packets and then after receiving a single ACK the source dumped another 900+ packets into the network. The tcpdump trace shows the rest of the 10 second test is consumed by re-transmitting everything from the beginning). I solved my problem by replacing the drivers/net/e100.c file with a version from the 2.6.12 release tree. I think it has something to do with the NAPI polling that was introduce in the 2.6.13 kernel tree.

Rich

At 12:33 PM 3/16/2006, Matt Crawford wrote:
Has NDT been known to be in error about reporting packet reordering?
Might there be bugs in web100 that mislead it? We're getting low
throughput on a 90ms RTT layer-2 path and NDT reports small amounts
of reordering. So far we are unable to find a switch to blame for it.

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



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