TTCP using TCP small servers Vs default port 5001

[r-godden]

I am

testing throughput on a circuit.

If i use TTCP at both ends with port 5001 i get a 52% util of link as result,

If i use TCP small servers and port 9 at one end i get 78% util as the result.

I have repeated the tests and get same results .

Any reason for the big difference ?

[Phillip Remaker]

Apparently, the discard port is a better performing TCP sink than the dedicated TTCP sink.  They are just different code bases.  TTCP is a much older code base (circa 1984!), which has long since replaced by newer systems like iperf.

It might be interesting to look at the TCP windows sizes sent back by each item.  Bottom line is that the TCP discard code is probably just better optimized than the ttcp sink.

[tslattery]

Phil is correct about differing code bases. Since you say that you're using small servers, it sounds like you're doing ttcp on Cisco boxes on each end of the link.  Some factors that will affect the performance:

• Window size, as mentioned by Phil, is the biggest factor, and will depend on the round trip delay of the circuit you're testing.
  
• The small servers code may run at, or close to, interrupt level, while ttcp probably runs as a user-level process. The user level process may get a lower level of service, causing it to be scheduled to run less frequently. If either the source or destination router is busy forwading packets (perhaps forwarding on a different interface than the one you're testing), the ttcp user level process will likely get a lower level of service than packet forwarding. The small servers code is probably running at a different level within the IOS system.
• Are either of the source or destination systems running the newer modular IOS? That would likely make a difference.
• The direction of your test A->B or B->A should also be tested to look for asymmetric patterns (may not be relevant to the problem you're reporting, but you don't mention testing in that direction).
• And as Phil mentions, the TTCP code may not be optimized like the TCP discard code. Hopefully, it isn't allocating a new buffer for each incoming packet. If it is, then that would account for a major performance penalty.

[Phillip Remaker]

In general, if you are looking for a serious stress test, you are better off using a dedicated test platform or at least specific, modern code dedicated to the job on a platform which can allocate enough resources to really soak the link.

I am not a expert on testing tools, but there is a lot of stuff out there.