On Sun, Dec 4, 2016 at 9:13 AM, Dave Taht <dave.taht@gmail.com> wrote:
On Sun, Dec 4, 2016 at 5:40 AM, Rich Brown <richb.hanover@gmail.com> wrote:
> As I browse the web, I see several sets of performance measurement using either netperf or iperf, and never know if either offers an advantage.
>
> I know Flent uses netperf by default: what are the reason(s) for selecting it? Thanks

*netperf
+supports multiple tests in parallel on the same server

 
* iperf
+ More widely available
Sort of...  Given the variants, less so.  But iperf3 is coded to be pretty portable, and so it's pretty widely available

It has a pretty good JSON format for the client results, but the server results are returned in plain text.  And it doesn't report anything finer-grained than 100ms.

- I have generally not trusted the results published either - but
aaron finding that major bug in iperf's udp measurements explains a
LOT of that. I think.

I've found something else with it, that I need to write up:  with UDP the application-layer pacing and the fq socket pacing cause it to report a lot of invalid packet loss.  The application pacing is focused on layer-6 good-put.  But the fq pacing appears to be enforcing wire-rates (or calculated ethernet rates), and so with small packets (64 byte payloads), it's typical to see 40% packet loss (as the fq layer discards the UDP frames to cut say 100Mbps of application layer down to 100Mpbs of wire rate).  I need to actually do the tcpdump analysis of that and get it written up.
 
- Has, like, 3-8 non-interoperable versions.
- is available in java, for example

there *might* be an iperf version worth adopting but I have no idea
which one it would be.

Part of the issue with iPerf is that there are two main variants:  iperf and iperf3.  iperf3 is currently maintained by the ESNET folks, and their use-case is wildly different from ours:

- Very high bandwidth (>=100Gbps)
- Latency insensitive (long-running bulk data transfers)
- private networks (jumbo frames are an assumed use)

I'm also happy to take the fork of it that I have (https://github.com/woody77/iperf) and make that tuned for our uses.  There are certain aspects that I wouldn't want to dive into changing at the moment (like the single-threaded nature of the server).  But I can easily bang on the corners and get it's defaults better suited for our uses, and make it behave better in the face of running without offloads.  On my test boxes, it starts to get I/O limited around 4-5 Gbps when using 1400-byte UDP payloads.  With TCP and TSO, it's merrily runs at 9.4Gbps of good-put over a 10Gbps NIC.  

But the application and kernel-level "packet rate" is really quite different at that point.  By default it's dropping 128KB blocks into the TCP send buffer, and letting the TCP stack and offloads do their thing.

At the higher-performance end of things, I think it would benefit from using sendmmsg()/recvmmsg() on platforms that support them.  I think that would let it better work with fq pacing at rates of 5Gbps and up.

I started speccing out a flent specific netperf/iperf replacement
*years* ago,  (twd), but the enormous amount of money/effort required
to do it right caused me to dump the project. Also, at the time
(because of the need for reliable high speed measurements AND for
measurements on obscure, weak, cpus) my preferred language was going
to be C, and that too raised the time/money metric into the
stratosphere.

iperf3's internal structure might be useful for bootstrapping a project like that.  It already has all the application logic infrastructure, and has a central timer heap (which can be made more efficient/exact), and it already has a notion of different kinds of workloads.  It wouldn't be too much work to make its tests more modular.

-Aaron