Re: Some questions about Iperf
Daniel;
Here's what I would do.
1) Run Iperf in loopback mode to determine how fast the memory subsystem
is. To run in loopback mode, start an Iperf server (iperf -sD) process on
one of your machines. Then start the client process on the *same* machine
pointing to the loopback address (iperf -c 127.0.0.1). Observe the
results. Monitor or capture the CPU utilization numbers.
2) run Iperf in 1 direction only. Start the server process on one machine
and the client on the other. Observe the results and monitor or capture
the CPU and network I/O utilization numbers.
3) run Iperf in the opposite direction. Reverse server and client
processes on the above machines. Observe the results and monitor or
capture the utilization numbers.
4) run Iperf in both directions. Rerun your previous test and monitor or
capture the utilization numbers.
*) if possible, install the Web100 (http://www.web100.org) code on the
client machine. This code will allow you to determine what TCP thinks is
happening during these tests. Note: the Web100 code only runs on Linux
OS's at the present time.
Rich
At 03:09 PM 12/6/02 -0800, Daniel E. Spisak wrote:
>At 04:54 PM 12/6/2002 -0600, Kevin Gibbs wrote:
>> > I guess what I am asking is, is Iperf only accurate for
>> > unidirectional tests?
>>
>>No it should be accurate in the testing scenario that you propose. One
>>thing to worry about it IO subsystems. Your built in NICs may not be able
>>to push 200Mb of data.
>>
>>Kevin
>
>What kind of I/O subsystem do I need to have to sustain 200Mbit/s of data?
>
>The 32bit/33Mhz PCI bus can handle 132MB/s in theory.
>
>Let's say I lose 20% of that to protocol overhead (I'm generalizing, I
>know its not that great)
>
>That still leaves me with 105MB/s of bus throughput.
>
>One 10/100 onboard NIC running at 200Mbit/s is only needed 25MB/s of that
>bus throughput.
>
>So how is my I/O subsystem substandard here? I mean, the onboard NIC's are
>Intel chips, and I can't say that I've ever run into any kind of serious
>problems with them before.
>
>Has anyone else on here tried this particular kind of test setup before?
>I'm wide open to suggestions.
>
>-Daniel Spisak
>
>
>
------------------------------------
Richard A. Carlson e-mail: RACarlson --at-- anl.gov
Network Research Section phone: (630) 252-7289
Argonne National Laboratory fax: (630) 252-4021
9700 Cass Ave. S.
Argonne, IL 60439