Re: Question about test results


On Thu, 22 Apr 2004, Dan Watts wrote:
> 
> A theory I have is that one side starts transmitting and fills the
> outbound TCP/IP queue so that it takes a significant amount of time for
> the TCP/IP connection establishment for the second connection to
> complete.  This would explain why the UDP test shows a very nice
> symmetrical result but the TCP/IP test does not.
> 
> I've thought about modifying the code so that the transmit test delays
> for a few seconds before starting but I think I may still run into the
> same issue with ACK packets getting stuck behind so many transmitted
> packets.

Dan,
  Your theories and thoughts are pretty dead on. The best way to get 
around this problem is not actually found in the delay or some other way 
to manipulate two connections to be nicer to each other. A more ideal 
solution is to actually use ONE connection. This is a feature that many 
people has asked for (gets around NAT issues, etc) but we have not 
implemented at this time. Also UDP is a better way to measure raw 
throughput of a pipe. The different congestion control mechanisms can 
generate wide differences in values and depends to a much greater extent 
on the end nodes (TCP implementation, window sizes, etc) rather than the 
network itself. Though obviously customers may be more interested in what 
they can expect from a HTTP, FTP, SSH aka TCP transfer.

Kevin



Other Mailing lists | Author Index | Date Index | Subject Index | Thread Index