Thanks. I'm curious to how much the threading helps performance here if at all. That part of the design we inherited from iperf 2 when we took over maintenance of it. We tried to keep the overall code similar in order to not impact performance. At first we tried not to touch the traffic threads "fast paths" but over time found we really needed to. We also kept the output the same for script compatibility. For newer enhanced output, the -e flag is required.
We're considering adding CPU measurements but probably not directly rather via coordination using the python controller model. Another possibility is a hotelling multivariate statistic to monitor CPU, memory, latency and average throughput - but that's all for regression detections and probably only of interest to automated test houses. Do let me know if you are able to get some data. We have access to WiFi class devices and platforms and some 10G but no machines that support 100G. Bob On Thu, May 30, 2019 at 12:41 PM Jeffrey Lane <j...@canonical.com> wrote: > On Thu, May 30, 2019 at 3:23 PM Bob McMahon <bob.mcma...@broadcom.com> > wrote: > > > > hmm, I confused. Did you run multiple iperf 3 sessions or iperf 2 with > the -P 8,10 option or possibly both? Your previous response said the only > way to get this was with multiple iperf 3 sessions and didn't mention iperf > 2 nor the use of -P. > > multiple iperf3 instances, similar to what Chris Preimesberger > demonstrates in his video, only we use upwards of 10 threads. (it > works out to about 1 thread per 10Gb of bandwidth). > > You mentioned iperf2 would be interesting to try, I was just > commenting that we went with multiple iperf3 instances vs a single > "iperf2 -P" because iperf3 does some things we wanted (like CPU > measurements) but lacks true multithreading. The "use multiple > instances of iperf3" is a workaround to lack of proper > multi-threading. Also, note, I'm not an expert, we came to this by > way of a lot of internet reading and trial and error. > > > In theory, iperf 2 could outperform iperf 3 per the use of threads, e..g > separating the traffic from the accounting and reporting. I'm curious to > actual experimental results. > > Note: Iperf 2.0.13 is really required for this class of testing as > older iperf versions (e.g. 2.0.5) have performance related bugs. > > It may. If one of my machines frees up I may try this to see how it > works out. No promises though, I've already got too much work as it is > :( > > > > > Bob > > > > On Thu, May 30, 2019 at 11:49 AM Jeffrey Lane <j...@canonical.com> > wrote: > >> > >> For my needs (very simple testing) yes. We had to do that because > >> iperf3 doesn't multi-thread like iperf 2 did, unfortunately. > >> > >> On Thu, May 30, 2019 at 1:37 PM Bob McMahon <bob.mcma...@broadcom.com> > wrote: > >> > > >> > Is it just multiple threads? It might be interesting to try iperf > 2.0.13 and the -P 8 option. > >> > > >> > Bob > >> > > >> > On Thu, May 30, 2019 at 10:04 AM Jeffrey Lane <j...@canonical.com> > wrote: > >> >> > >> >> I've been working on this a bit and the only way to get it was to run > >> >> multiple iperf3 threads. To do this, you have to set up several (we > do > >> >> about 8 threads for 100Gb, possibly 10) on the target (listening to > >> >> different ports) and then run to client instances (one for each > port), > >> >> then aggregate the results for each, and that nets in the 92-97Gb/s > >> >> range overall. > >> >> > >> >> Additionally, in some cases tweaks are necessary (jumbo frames, some > >> >> kernel tweaks, driver tweaks, etc) but that's all case-by-case. > >> >> > >> >> And it is very much constrained by CPU and PCIe bandwidth. > >> >> > >> >> > >> >> On Thu, May 30, 2019 at 12:38 PM Chris Preimesberger < > ccpi...@gmail.com> wrote: > >> >> > > >> >> > I tried and got up to 87Gbps throughput. The results were CPU > bound. I want to build new i7 9900K PCs and re-test. Here's a video of my > attempt: > >> >> > > >> >> > https://youtu.be/uh2zvaaH0hc > >> >> > > >> >> > > >> >> > > >> >> > On Thu, May 30, 2019, 3:08 AM Ashwajit Bhoutkar < > bhout...@gmail.com> wrote: > >> >> >> > >> >> >> Hi, > >> >> >> > >> >> >> Just wanted to check whether it is possible to test the > throughput of 100G link using iPerf. > >> >> >> > >> >> >> > >> >> >> Thank You, > >> >> >> > >> >> >> Kind Regards, > >> >> >> Ashwajit > >> >> >> _______________________________________________ > >> >> >> Iperf-users mailing list > >> >> >> Iperf-users@lists.sourceforge.net > >> >> >> https://lists.sourceforge.net/lists/listinfo/iperf-users > >> >> > > >> >> > _______________________________________________ > >> >> > Iperf-users mailing list > >> >> > Iperf-users@lists.sourceforge.net > >> >> > https://lists.sourceforge.net/lists/listinfo/iperf-users > >> >> > >> >> > >> >> > >> >> -- > >> >> Jeff Lane > >> >> Engineering Manager > >> >> IHV/OEM Alliances and Server Certification > >> >> > >> >> "Entropy isn't what it used to be." > >> >> > >> >> > >> >> _______________________________________________ > >> >> Iperf-users mailing list > >> >> Iperf-users@lists.sourceforge.net > >> >> https://lists.sourceforge.net/lists/listinfo/iperf-users > >> > >> > >> > >> -- > >> Jeff Lane > >> Engineering Manager > >> IHV/OEM Alliances and Server Certification > >> > >> "Entropy isn't what it used to be." > > > > -- > Jeff Lane > Engineering Manager > IHV/OEM Alliances and Server Certification > > "Entropy isn't what it used to be." >
_______________________________________________ Iperf-users mailing list Iperf-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/iperf-users
