iperf to test bandwidth, what setting are best?

[Andy White]

Hello,

I have configured a few 3750G switches with trunks and I want to test the throughput.  In your experience is iperf the best tool?  If so what settings do you use?

On a quiet network would you expect to see 900-1gb throughput?

Thanks

[rsimoni]

Hi Andy,

among the non professional tools iperf is (one of) the best.

Test the link by using udp traffic only (not tcp) with different frame size.

About the link utilization it really depends on your network patterns.

In general having a gigalink approaching 90% of its bandwidth is not a good sign as you might easily see drops (due to burstyness of traffic).

Riccardo

[Andy White]

Thanks, do you use UDP for better results as the header is much smaller and no 3-way handshake?

Regards

[rsimoni]

Yes, but mainly because on switches you can see performance issues for bursty TCP applications. The reason is simple, if some packets get dropped because the buffer cannot contain it the side effect of TCP sliding window will kick in causing a reduction of the transmission window (trasmit rate slows down). Consequence of that is a reduced throughput and therefore an non optimal performance measurement.

This does not apply for udp instead.

Riccardo

[Peter Paluch]

Hi Riccardo,

Hey, it's been quite a time since we last met in a thread How are you?

Please allow me one question, or better say, a comment - and I am interesting in knowing what you think.

Testing the throughput of a switch, in my opinion, does not reveal much if only a single device attached to a switch generates the traffic load. The switching fabric of a switch has its capacity bigger by orders of magnitudes than any switchport on this switch. Loading a single ingress port even to its top is not going to put the switching matrix under any reasonable stress (except for broadcasting or multicasting, and event then I tend to believe that the load incurred on the switching matrix is negligible under this setup).

So to put the the switch under real stress would mean to have the switch fully connected (all switchports being connected to devices with traffic generators or traffic generator sinks) and have, by a rule of thumb, at least 50% of these devices send the traffic.

Am I thinking correct, or is this a wrong approach?

Thank you!

Best regards,

Peter

[rsimoni]

Hi Peter,

you are time, I have been absent from CSC for quite long... lots of travelling

anyway what you write is correct, testing a switch with just one ingress and one egress port connected does not stress it at all.

I am not sure whether Andy wants to just measure performance of  his switches rather than the end to end network though.

Riccardo

[adam dirkmaat]

Let me add something to this discussion... how can you simulate 50% load on a circuit to then test/document latency at 50% bandwidth utilization?  This is for capacity planning purposes.