Custom thread pool support
Ron Pressler
ron.pressler at oracle.com
Tue Jun 7 10:38:20 UTC 2022
Hi.
Yes, the plan is to ultimately allow custom schedulers, but we want to first let
the ecosystem learn about virtual threads and their uses with the default
scheduler.
I cannot study your code at this time, but custom schedulers will not be able to
break any specified invariants for threads, although they would be able to break
some unspecified expectations around liveness, as well as add scheduler-specific
invariants (e.g. that all memory by a set of virtual threads scheduled by the scheduler
are fully ordered with respect to all reads by the same set of threads) that don’t
necessarily hold for all threads in Java.
— Ron
> On 7 Jun 2022, at 10:52, Vladimir Ogorodnikov <vog at activeviam.com> wrote:
>
> Hello,
>
> My name is Vladimir Ogorodnikov, I am an R&D intern at ActiveViam. In my
> internship, I explore the possibilities of the Loom project.
>
> One of the tasks I solve is NUMA (Non-Uniform Memory Access) support. The
> essence of this architecture is that each processor has its own block of
> memory (processors and memory blocks connected to each other are called
> NUMA nodes). If the application is organized in such a way that each
> processor works with its own part of the data, then the performance grows
> by a multiple of the number of processors (or memory blocks). Processors
> can also access remote memory blocks, but this requires additional time
> spent on data transfer.
>
> An important part of NUMA support is the ability to specify for each thread
> which processors it can run on. To do this (on Linux) one need to make a
> native call to numa_run_on_node(). On other OSes, the procedure may be
> different. This cannot be done from the code of a virtual thread: firstly,
> virtual threads are invisible for the OS, and this call will change the
> properties of the carrier thread, and secondly, the virtual thread can be
> moved to another carrier thread at any time.
>
> Thus, full support for NUMA requires the ability to create several
> different pools of carrier threads with the ability to configure them. This
> is not possible in the current version of the API. I prepared a demo
> project in which the required functionality is implemented using reflection
> and changing the visibility of methods. In addition, I prepared a virtual
> machine image for QEMU on which you can run this project. The project is
> available on GitHub:
> https://github.com/activeviam/NumaAwareVirtualThreadsDemo. The files are
> available at:
> https://drive.google.com/drive/folders/1v7xn8NEELfzR9TCn5hrV52n03hfOHeAh?usp=sharing
> .
>
> I would like to ask if there are any plans to add support for this
> functionality in Loom, and if my code breaks any internal invariants
> related to virtual threads. Thanks in advance.
>
> Sincerely,
>
>
> Vladimir Ogorodnikov
> *R&D Intern*
> [image: Mobile] 0778300738 [image: ActiveViam]
> <https://www.activeviam.com/>
> [image: Linkedin] <https://www.linkedin.com/company/activeviam/> [image:
> YouTube] <https://www.youtube.com/c/ActiveViam/videos> [image: Blog]
> <https://www.activeviam.com/blog/> [image: Twitter]
> <https://twitter.com/activeviam>
> 46 rue de l'Arbre Sec, 75001 Paris
> <https://maps.google.com/?q=46+rue+de+l+Arbre+Sec,+75001+Paris,+France>
> www.activeviam.com
More information about the loom-dev
mailing list