Closed evaleev closed 10 months ago
Interesting ... what's driving this?
Interesting ... what's driving this?
larger multirank TA jobs (with hundreds of thousands of expression elements, involving thousands of tiles each) are getting stuck stochastically (but with near 100% probability) due to lack of progress. I thought maybe a lifetime issue (pulling from temporary objects, e.g. subexpressions), but does not seem to be that. Expression evaluation actually pushes data up, which makes lifetime management easier.
@robertjharrison I don't think I made any perf-related changes here, but if you see something amiss let me know.
Is there an issue with ordering or tasks internally forcing futures? Or use of send instead of task?
Grasping at straws in the absence of info.
On Thu, Dec 28, 2023, 22:09 Eduard Valeyev @.***> wrote:
Interesting ... what's driving this?
larger jobs (with hundreds of thousands of expression elements, involving thousands of tiles each) are getting stuck stochastically (but with near 100% probability) due to lack of progress. I thought maybe a lifetime issue (pulling from temporary objects, e.g. subexpressions), but does not seem to be that. Expression evaluation actually pushes data up, which makes lifetime management easier. TA expression cleanup does not use global fencing, only local info, and push makes it possible.
— Reply to this email directly, view it on GitHub https://github.com/m-a-d-n-e-s-s/madness/pull/512#issuecomment-1871695892, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABZSAPOUOLFLTDJW4KN6HKTYLYYADAVCNFSM6AAAAABBE5DZI6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNZRGY4TKOBZGI . You are receiving this because you commented.Message ID: @.***>
World::id_to_ptr()
return an optional; ifNDEBUG
is not defined then its value for deregistered IDs is nullptr, this can be used to detect use of unregistered/deleted objects.