[omniORB] strange TCP/IP effect, anyone an explanation ?

[Sascha Kümmel]

Hello,

I'm not an expert in omniORB but the second effect you described looks like
a well-known problem of request-response communication over some TCP/IP
stacks. We observed the same with older versions of DCE or Microsoft RPCs
some years ago.  

>
>Effect 2:
>
>In a second test we measured the distribution of the response times of a
>remote upcall against increasing parameter size. 
>When the message size is bigger than 4 KBytes, the response time of the
>remote upcall increases by factor 1000 (see picture rate2.gif).
>We believe that the cause of this effect is situated in the TCP/IP stack of
>solaris, but we can't find a resonable explanation.
>Has anyone encountered similar effects ?
>

The effect you observed is called the 'silly window syndrome'. If the size
of a request is a multiple of 4000-4100 the request will be delayed by
approx. 200 ms (a timeout value in the stack). In DCE 1.1 this problem with
some TCP/IP stacks was fixed by increasing such a message with some padding
bytes. This waste bandwidth but is a 'compatible' solution.

Regards,

Sascha Kuemmel