Re: [boost] interest in an asynchronous linear algebra library?

Subject: Re: [boost] interest in an asynchronous linear algebra library?
From: Oswin Krause (Oswin.Krause_at_[hidden])
Date: 2015-11-19 13:13:13

Hi,

> This sounds like a specific solution to a general need for schedulers.

This proposal can work on top of any scheduler that allows for 
dependency graphs. You could see this library as a specific application 
of scheduling linear algebra tasks, because we know exactly how the 
dependencies are and how to order stuff. Therefore, we can completely 
hide the scheduling.

> Have you looked at the current proposals from the various sg groups
> such as sg1? Do they not help to solve your problem? Now is the time
> to speak up if they don't. Go find Michael Wong or Hans Boehm.

I am not sure about the current state of the proposals, I only remember 
a recent discussion about executors where I raised concerns about 
possible deadlocks and missing guarantees. Anyways, I would not know 
where to start looking for the current state of the proposals.

I found this here:
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4414.pdf

and I doubt one could implement dependency graphs with that proposal as 
the executor does not return references to scheduled tasks or allows for 
explicit dependencies(i can not use futures as I want operations as much 
as possible to use existing objects and not generate new objects as this 
can exhaust memory pretty easily). There is also no wording about "what 
happens when a work package adds more work packages to the executor and 
waits for them", which makes any thread_pool proposal worthless as 
blocking worker threads might result in deadlocks. I think there is 
further no hope that we get unified dependency handling which allows for 
tasks handled by different executors/schedulers (a CPU and GPU scheduler 
fx).

Best,
Oswin