From document about I_MPI_ADJUST Family Environment Variables, there is a description:
algid >=0 The default value of zero selects the optimized default settings.
But when I find the alg_id of I_MPI_ADJUST_ALLREDUCE with impi_info, I found that the default value is -1.
What does -1 mean here? What't the default alg for allreduce? Or is it dynamic adjusted?
Thanks for reaching out to us.
You can see the range I_MPI_ADJUST_ALLREDUCE can take using the impi_info - v I_MPI_ADJUST_ALLREDUCE command.
It has been mentioned that it can take values from 0-24 where 0 as mentioned is the default value.
If 0 is selected the MPI uses the out-of-the-box tuning data to select the best algorithm.
Regarding the -1 as the default value in the impi_info, we will get back to you.
Thanks & Regards
Before we resolve the default alg id=-1, I got another question during the test.
As the docuemnt mentioned, I_MPI_ADJUST_ALLREDUCE has 12 optional algorithms. The version of the document is lastest 2021.2(Last Updated: 04/06/2021).
But the output of impi_info -v I_MPI_ADJUST_ALLTOALL shows that algid range from 0-26. The number is much greater than that in document.
Where could I find the name of all the 26 algorithm?
Thanks for bringing these topics up here.
There is a mismatch in the current version of online documentation and the output of impi_info in regard to the default values of I_MPI_ADJUST<opname> variables. We are working on correcting this. Please note that the default value of I_MPI_ADJUST<opname> variables is -1, as indicated by the impi_info utility.
In regard to your other query on the number of algorithms in I_MPI_ADJUST_ALLREDUCE please note that the online documentation shows the algorithms, while impi_info shows the number of algorithm presets in the range field.
Impi_info shows the number of algorithm presets in the range field, but the documentation only shows part of the algorithms with number----alg_name.
When will the latest online documentation be released?
Yes, that is correct.
We expect to fix the documentation bugs along with the release of the next version of Intel MPI Library.
Just wanted to check if you had any further questions before closing this thread?