Let IMT influence TThreadedObject #1019
Conversation
If IMT is enabled, TThreadedOject should allocate as many slots as threads the pool has been set with.
| fModel.reset(Internal::TThreadedObjectUtils::Detacher<T>::Detach(new T(std::forward<ARGS>(args)...))); | ||
|
|
||
| if(ROOT::IsImplicitMTEnabled()) { | ||
| TThreadedObject<T>::fgMaxSlots = ROOT::GetImplicitMTPoolSize(); |
There was a problem hiding this comment.
Why change the value for each TThreadObject creation?
Also, in order for this line (and the next) to be thread safe fgMaxSlots needs to be an std::atomic. [Otherwise, per the standard, the behavior when the reading (line 167) and the writing (line 164) happen at the 'same time' is undefined]
There was a problem hiding this comment.
Good point. Will look into it 🤔
There was a problem hiding this comment.
@pcanal doesn't everything go bonkers anyway if users call EnableImplicitMT in a thread while they are using implicit-mt-aware features in other threads concurrently?
In other words, shouldn't EnableImplicitMT be called strictly before doing anything that checks whether implicit-mt is enabled or not?
|
Starting build on |
|
Does fgMaxSlots have to be static? I can't figure out the reason apart from wanting to keep the slot size heterogeneous at any point of the program. I would need it to be local for dependency on IMT behaviours to make sense. |
|
@dpiparo What do we make of this? It needs a refresh, but is this the right approach performance wise? |
1 similar comment
|
@dpiparo what do you want to do with this one ? |
|
IIRC, not relevant anymore. Closing. |
7 participants
If IMT is enabled, TThreadedObject should allocate as many slots as threads the pool has been set with.
This PR will be updated with another default value for fgMaxSlots once PR #1018 has been agreed upon and merged.