[netperf-talk] Confusing result of the throughput in wireless network with different UDP packet length.

[Rick Jones]

Andrew Gallatin wrote:
> Peng Wang wrote:
> 
> 
>>The benchmark platform I use contains two buffalo wireless g router. One
>>router is the sender, and one is the receiver. The operating system in
>>them is Linux with 2.4.7 version kernel.
>>
>>The result of the benchmark confuses me. From the following figure we
>>can see that the throughput is highest when the UDP packet size is about
>>20k bytes which is much bigger than the MTU. I guess it is because of
>>the buffer in the system, but I am not quite sure. Do you see it before?
>>Could you explain it?
> 
> 
> I'm not at all familiar with performance characteristics of
> wireless networks, but on fast (10GbE) wired networks I
> sometimes see similar behavior when the host has expensive
> system calls.   The larger the message size, then the more data
> is written on the socket using a single system call.

It doesn't necessarily have to be an "expensive" system call either.  When 
sending 1000 bytes at a time that is 20 system calls and 20 calls through the UDP 
code into IP and then 20 trips through the driver and onto the "wire."  With 20K 
bytes at a time, that is one system call, one call through UDP into IP and then 
20000/MTU trips through the driver.  If MTU is larger than 1000 bytes then that 
will be rather less than 20 trips through the driver - perhaps only about 15 or 16.

> So, a few questions:
> 
> 1) Does a test via a 100Mb or 1Gb connection between the routers show
>    the same characteristic?   How about a local (loopback) test
>    from a machine to itself?
> 
> 2) Have you measured the system call overhead using something like
>    lmbench?

I would ask:

3) Where are the netperf-measured service demands? (ie add CPU utilization 
options to your test)

4) Does the graph show reported send throughput, or actual receive throughput?

5) Was send throughput == receive throughput?

rick jones