[netperf-talk] Re: Fwd: reg netperf executions on W2K3 64bit platform

[Rick Jones]

sumit panchasara wrote:
> Hi
> 
> I have observed following behavior on netperf and netserver
> side respectively for following commands:
> 
> Server: netserver.exe -L 192.168.1.10 -p 9375
> Client: netperf.exe -H 192.168.1.11 -p 9375 -t TCP_STREAM -l 6 -f M --
> -m 1048576 -M 1048576 -s 5376000
> 
> Netperf: (client)
> 1. we have got winsock direct DLL main's process detach case
> 2. that will call all resource cleanup routines, and lead to connection 
> teardown
> 
> Netserver: (server)
> 1. windows socket switch has somehow called CloseSocket routine of WSD 
> code.
> 2. That ultimately results into socket teardown, and iWarp resource release
> 3. so all the connections are getting teardown here too for this socket.
> 
> Is this abnormal termination is due to dll detach call to wsd?
> If yes what could be reason which causes this to happen on WSD path?
> why server side applicaiton calls closesocket which maps to winsock
> direct closesocket call and does teardown?
> 
> The same is working on non-accelerated path, i.e. which goes
> through TCP/IP stack and uses NIC oriented functionality.
> 
> Also I tried other tools like, NTttcp and chariot which doesn't shows
> such behavior and works fine for 8192 data size, 1 stream
> unidirectional connection properly...

Which connection is this again?  The netserver will "fork" (spawn a new process) 
after the control connection is accepted.  Does the DLL "handle" that well?  I 
ask because NTttcp does not (IIRC) "fork."

If there is a way to tell the DLL to leave control connections alone (eg the 
port 9375 in your example above - that is the port used for the control 
connection, not the data connection) it might be a good thing.  The "data" 
connection will be created _after_ the child netserver process is created, so no 
issues about inheritance of the endpoint across process creation.

happy benchmarking,

rick jones