Disabling threading completely -- sequential mode

[mike-jn]

Hi,

I would like to disable threading completely since my application runs in single-threaded mode, pinned to a single core. Basically, I want the algorithm.compute() function to run in the caller thread and not dispatch work to any other thread.

I understand that MKL, for example, has a sequential mode library that is made specifically for usecases like this, see: https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-linux/2025-0/linking-with-threading-libraries.html

How can I achieve the same with oneDAL/DAAL?

Thanks!

cc @napetrov

[napetrov]

You can only set thread to 1, there is no separate sequential library for oneDAL. https://uxlfoundation.github.io/oneDAL/daal/services/managing-computational-environment.html you can set it with library API or by global controls in TBB

@Vika-F @david-cortes-intel

[mike-jn]

Thank you for such a quick response!

What does setting it to 1 mean in practice? Will there be an additional thread created by the library?
Note that my app is pinned to a core and no other thread can run there, so I need to better understand how threading works with oneDAL.

Additionally, does it use MKL under the hood, or is it self-contained? (note: I only use the GBT prediction feature). I am using the builds and it doesn't seem that the dynamic library is linked to MKL on a quick inspection (maybe it loads it manually?).

[Vika-F]

oneDAL relies on oneTBB as a parallel runtime. Setting 1 thread in oneDAL will result in re-initializing oneTBB's scheduler with only 1 worker thread using tbb::global_control(tbb::global_control::max_allowed_parallelism, 1).
I have contacted oneTBB team to provide the exact answer about how many TBB threads will work on the system in this case and will return to you later with the answer.

Also yes, from version 2025.0 oneDAL uses oneMKL under the hood (and oneMKL also uses oneTBB as a threading layer in this case). So, the threading setting would be propagated to oneMKL as well.

And here is the code to limit the number of threads in oneDAL:

#include "daal.h"
#include "service.h"
...
constexpr size_t nThreads = 1;
daal::services::Environment::getInstance()->setNumberOfThreads(nThreads);

link to the example

[mike-jn]

Thank you very much for the info. Using additional threads is not good in our case since we have a low-latency single-threaded app. We would need something like MKL's sequential mode for this to run properly in our runtime environment.

How hard do you think it would be to modify the existing code so that compute() runs in the caller's thread?

Note: I am currently on 2024.* and cannot upgrade to 2025 due to the requirement of a newer glibc version of the latter.

[napetrov]

we still was using custom MKL build prior to 2025.0 release. so you should be getting similar behavior

On complexity - this would depend on ammount of algorithms and specific details there. But in the past there was sequential lib of oneDAL, but it was not used so we dropped it. So technically this is possible especially if you would be looking for single algo

[napetrov]

@mike-jn can you share more details on your usecase? How you train your models? would be happy to talk in email/meeting if you prefer not to share publicly.

with setting to 1 we would create single thread. The only difference from true "sequential" library is that we still would initialize threading layer which might bring some overhead.

As for MKL - we do control over it's usage and it shouldn't thread more than setting at oneDAL level. In tree based algos we use RNG portion from MKL

@Vika-F can comment more on threading layer

[mike-jn]

@napetrov email would be great, can I drop a msg to the email in your profile?

I am currently still testing out the library and have a bunch more questions on using it optimally based on our use-case.

[napetrov]

yes, sure.