hkaiser changed the topic of #ste||ar to: The topic is 'STE||AR: Systems Technology, Emergent Parallelism, and Algorithm Research | stellar.cct.lsu.edu | HPX: A cure for performance impaired parallel applications | github.com/STEllAR-GROUP/hpx | Buildbot: http://rostam.cct.lsu.edu/ | Log: http://irclog.cct.lsu.edu/
<simbergm>
I'll add a note saying that your scheduler is still experimental
<simbergm>
so unless you already know what's wrong no need to fix it right now
<jbjnr>
I have no idea. I have a soak test running here - repeating the cross-pool-injection test 1000 times and it's been going for a while with not a single fault
<jbjnr>
If I can reproduce it, I can fix it,
<jbjnr>
The reason I created the test is because injecting tasks across thread pools is "risky" and I wanted to catch exactly this kind of thing
<jbjnr>
it has to be a thread numbering/indexing issue
<jbjnr>
maybe it only happens on multi numa nodes?
<jbjnr>
Unfortunately, rostan buildbot doesn't load fr me, so I can't see the errors
<jbjnr>
simbergm: "I think that'll start the task before the future is ready" I'm not sure I understand you (Comment earlier about async_customization example)
<simbergm>
hmm, what do you get? blank page?
<simbergm>
what I was trying to say is that the continuation task is started before the future is ready
<simbergm>
i.e. what you're doing there is equivalent to future f = async(fun1); future g = async(fun2, f);
<simbergm>
*the continuation task is created before the future is ready
<jbjnr>
are you sure you pasted the right link?
<jbjnr>
which lines are you worried about?
<jbjnr>
oh I see
<jbjnr>
but thae then_execute doesn't get called at all until the future is ready anyway
<jbjnr>
1000 runs of cross poll test finished without error
<jbjnr>
grrrrr.
<simbergm>
I don't think then_execute gets called only when the future is ready
<simbergm>
make_continuation* attaches a continuation that is run when the future is ready but I think then_execute is called right away
<simbergm>
for the cross_pool_injection test it seems like the fails quite evenly across the compilers (maybe 50% from 1 run...) but always in release mode
<jbjnr>
rostam loaded. I'm looking at errors now.
<jbjnr>
then_execute returns a future, but the task that makes that future ready is only scheduled in the continuation. (Or so I believed - however, that async test might well be wrong - if that is the case, then so are the guided executor and the limiting executor - and they should be fixed). I guess I'm using the old API .... and the stuff needs to be cleeaned up?
<jbjnr>
I guess it's the call to apply that's the problem
<simbergm>
I think the guided executor actually does it correctly, it's just that example (AFAICT now)
<jbjnr>
oh. then I messed up in some copy paste. The async exec might be out of date then.
<simbergm>
yep, I think it's just a mistake in that one
<jbjnr>
that one looks more sensible
<jbjnr>
the guided pool exec does the same
<jbjnr>
it's just hidden by the rather complicated indirection through the helper object
<jbjnr>
they could all be wrong
<simbergm>
indeed
<simbergm>
it also uses dataflow which just takes care of that
<jbjnr>
though the tests would not work if the continuation was being run before the future - I'd better check the details
<jbjnr>
well the guided pool is ok
<simbergm>
it'll still work because future::get blocks
<jbjnr>
it uses an internal dataflow
<simbergm>
the semantics are correct, it's just not the most efficient implementation
<jbjnr>
which bit is inefficiet - the routing throught the future factory?
<simbergm>
just the potential blocking on getting the future in the continuation
<jbjnr>
I'll upgrade them to use the same semantics as the parallel_executor
<simbergm>
I mean the wrong example, not your guided executor
<jbjnr>
NB. the guided pool executor is supposed to block in the continuation. It delays the scheduling util all the futures are ready so that it can do introspection
<jbjnr>
hence the use of dataflow
<jbjnr>
but the spawning of the continuation might be inefficient - using the future factory
<jbjnr>
(well actually, all continuations block and delay scheduling, but this adds the introspection/capture step)
<jbjnr>
bbiab - coffee
<jbjnr>
simbergm: I am not sure what you are taking about when you say "I mean the wrong example, not your guided executor" - could you please explain which test/executor you're unhappy with in terms of efficiency
<jbjnr>
simbergm: you are correct. The .then implementation is wrong in async_customization. I have made the fix and improved the test to help me see it.
<jbjnr>
good find btw
simbergm has quit [Ping timeout: 240 seconds]
hkaiser has quit [Ping timeout: 248 seconds]
rori has joined #ste||ar
aserio has joined #ste||ar
hkaiser has joined #ste||ar
<hkaiser>
aserio: yt?
<aserio>
yep
<hkaiser>
hey Adrian
<aserio>
Good morning
<hkaiser>
could I join the stellar group meeting remotely?
<aserio>
Sure, let me set that up
<hkaiser>
thanks!
<aserio>
How is adjusting back to central time going?
<hkaiser>
aserio: better than the other way around ;-)
<hkaiser>
aserio: I hope to have some time to chat with you tomorrow!
<aserio>
hkaiser: oh do we have a lot to talk about!
<hkaiser>
indeed
<aserio>
as well as lots of meeting with students planned
<hkaiser>
uhh ohh
<aserio>
lol
<hkaiser>
aserio: see pm, pls
<aserio>
um... how did git not pull in the hpx/libs folder?
<aserio>
huh, apparently you have to pull again after checking out a remote branch
<aserio>
the more you know
<diehlpk_work>
hkaiser, We need to finsih the book proposal.
<diehlpk_work>
Do you have time to work on it today?
<hkaiser>
diehlpk_work: nod, I'll be at cct tomorrow
<hkaiser>
what should I do today?
_bibek_ has quit [Quit: Konversation terminated!]
simbergm has joined #ste||ar
bibek has joined #ste||ar
<diehlpk_work>
hkaiser, Please list names and affiliations of reviewers and specify whether they are conference program committee members or external reviewers?
<diehlpk_work>
Question 3.3.
<diehlpk_work>
And we need answers for 2.8 to 2.13
diehlpk_work has quit [Remote host closed the connection]
diehlpk_work has joined #ste||ar
rori has quit [Ping timeout: 248 seconds]
<hkaiser>
aserio: could you send the link for the meeting, pls?