Unclear meaning of options of mpitune

[oleotiger]

From the mpitune document , there are many options and I think I don't understand some of them. Can someone help me understand them better?

1. In the description of  -tree,

   The second last group contains variables to be optimized (*, denotes all the variables not present the other variable groups).​

   I think the variable to be optimized is I_MPI_ADJUST_BCAST. Why is there * instead?
2. -tree-ignore, -tree-intervals, -tree-tolerance. There 3 options have too simple description. Can you give me more details on how to use them with examples?

[SantoshY_Intel]

Hi,

Thanks for reaching out to us.

We are working on it and get back to you soon with an update.

Meanwhile, you can also refer to autotuner(Autotuning (intel.com)) which is more efficient and faster than mpitune.

Thanks & Regards,

Santosh

[oleotiger]

1. Before I choose mpitune to tune my applicaiton, I have read the document. I thought the invoking chain was: mpitune_fast -->mpitune-->autotuning.  It there invoking relation among them?
2. The output for mpitune is a json file which is readable. I can easily understand the meaning and compare the algorithm choice between that I did with OSU microbenchmark.  But the output of mpitune_fast and autotuning is a binary file. How to read or translate it to a readable file?
3. According to your reply, autotuning walks far more forward. Is there any document that describes the method how autotuning improve the performance?  I tried to test the performance of each algorithm for each message size for a specified collective operation with microbenchmark OSU. And according to the result I conclude a rule to a json file applied with I_MPI_TUNING. However, there is little improvement in performance...So I think there may be other methods or tips  that intel applied. I'll appreciate it if you could share the idea or related documents/papers.

[SantoshY_Intel · Accepted solution]

Hi,

>>" I thought the invoking chain was: mpitune_fast -->mpitune-->autotuning. Is there invoking relation among them?"

No, we have 4 tuning utilities which allow us to do cluster-wide & application-specific tuning:

1.  MPItune
2.  Fast tuner
3.  Autotuner
4.  mpitune_fast

For more information please refer to the below link and go through the article "Making HPC Clusters More Efficient using intel MPI Library" on  [page No: 53]:

 https://software.intel.com/content/www/us/en/develop/download/parallel-universe-magazine-issue-41-july-2020.html

>>" But the output of mpitune_fast and autotuning is a binary file. How to read or translate it to a readable file?"

 We can not convert it into a readable format. We just need to use it as an input for subsequent MPI commands and it takes care of it.

>>"I'll appreciate it if you could share the idea or related documents/papers."

 In the below link Intel demonstrated performance gains on 6 real-world HPC applications. please refer to the link below and go through the article "Boosting The Performance Of HPC Cluster Workloads Using Intel MPI Library" on [page No: 43]:

 https://software.intel.com/content/www/us/en/develop/download/parallel-universe-magazine-issue-42-october-2020.html

Thanks & Regards,

Santosh

[SantoshY_Intel]

Hi,

We haven't heard back from you. Is your issue resolved? Please let us know if you need anything else that we can help you out with.

Thanks & Regards,

Santosh

[oleotiger]

I'm working on it and thank you for your answer. It really helps.

If I have other questions, I'll post another topic.

[SantoshY_Intel]

Hi,

Thanks for accepting as a solution.

As your issue has been resolved, we will no longer respond to this thread. If you require additional assistance from Intel, please start a new thread. Any further interaction in this thread will be considered community only.

Have a Good day!

Thanks & Regards

Santosh