[netperf-talk] Beginner questions about netperf

[Rick Jones]

George Valzcig wrote:
> Hello,
> I am a novice at networking - trying to use netperf to understand the 
> basics of networking. I have a few questions about netperf.
> 
> 1) The _RR tests are implemented.
> 
> In TCP_RR (if my understanding of TCP is right), when a message is sent 
> on the data socket, it will ultimately be received by the netserver on 
> the remote end. It may take either one or many recvs() to get the 
> complete message - which is why the recv() spins in a loop till the 
> expected number of bytes is got. Then the netserver send a response to 
> the netperf client. So there is a one-one relation between every sent 
> message from the netperf client and every completely received message at 
> the netserver server.
> 
> In UDP_RR, lets say a message was sent - but it failed to reach the 
> server end. The netperf client will not know this. In theis case, wont 
> the server be waiting for-ever without sending the response(as it cannot 
> send a response without receiving the lost message) - which in turn will 
> dead-lock the netperf client (as the client cannot go to the next send 
> without receiving the response)?

That is correct.  The initial, and continuing philosophy behind netperf is that 
it is first and formost a stack performance benchmark, so it does not attempt to 
retransmit lost requests or responses in a UDP_RR test.

The deadlock will be broken when the test timer expires - the potential to lose 
a UDP datagram is why UDP_RR must always be a timed test and cannot be a 
transaction count test like TCP_RR can.

> 2) What is the reason for dirtying the send and recv buffers (in 
> access_buffer()) - is it to simulate "using" the data that is being sent 
> and received - like a real application?

More or less - it was used mostly in the days when HP-UX 9.X had copy avoidance:

ftp://ftp.cup.hp.com/dist/networking/briefs/copyavoid.pdf

but does get used from time to time today.

> 
> I humbly apologize if any of these questions have already been psted on 
> the list - that brings me to my third question
> 
> 3) Reading past archives
> Is there a way to read the past archives without searching through every 
> gzip'd file.

No idea :)  Sometimes it is nothing short of a minor miracle the archives exist 
at all :)  If folks are familiar with "better" ways to deal with email list 
archiving I'd be happy to hear about it.

happy benchmarking,

rick jones

> 
> Thank you,
> George
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> netperf-talk mailing list
> netperf-talk at netperf.org
> http://www.netperf.org/cgi-bin/mailman/listinfo/netperf-talk