[--rate_limit] A stable relative error (~7.5%) observed about perftest

[HassanKhadour]

Our pleasure!
Please check it if you can
1ef49c4

[doctormin]

[HassanKhadour]

Hi doctormin, thanks for the info!
Can you please help on how to reproduce it ?

[doctormin]

Hi Hassan! I am afraid that I can not share the .py script immediately because this test was done by me when I was working for an enterprise (I don't have access to it any more). But I can explain more about what happened in my test script:

• In a for loop, the arg to --rate_limit was set in range(5, 5750, 5) and ib_send_bw (or other command) was then executed each time
  • then the "actual rate achieved" was got from the terminal output of perftest everytime it finished.

So the script can collect multiple <input: rate_limit>, <output: actual_rate> data pairs. That's how I draw the diagram above.

This relative error was stable and reproducible in my experience when I was working on it. Could it be that the input (args --rate_limit) is throughput and the output of perftest's terminal report is goodput?

[HassanKhadour]

Hi, perftest output is goodput, I'm trying to reproduce this issue but I can't get those relative errors, I can get barely 1.5%~ at maximum

[doctormin]

I see, thanks for the information. Maybe the packet size setting will influence the gap between throughput and goodput (but I am not so sure). In my previous test (~7.5% result), I select the very packet size which can make the throughput to be MAX in my testbed (by carrying out -a test first), but I didn't remember the exact value of it. I am sure that I was using the same packet size setting for all my data points in the diagram above.

[HassanKhadour]

Hi @doctormin, We investigated the Issue and soon a fix will be merged.
The problem is in some formula we were dividing integers and therefore the fractions were ignored, now we changed the types of those variables to be "double".
After the Fix the results will be much more accurate.

[doctormin]

Happy to hear about that :)
Thank you guys for a great job

[HassanKhadour]

Hi @doctormin,
Can you please update us if u checked it?

[doctormin]

Hi @HassanKhadour,
I am sorry that I can not check this issue in near future.
Because this issue was discovered when I was an intern in an HPC-related enterprise, and currently I don't have access to an IB cluster any more.

[HassanKhadour]

No worries, thanks for reporting the Issue, Closing