hkaiser changed the topic of #ste||ar to: STE||AR: Systems Technology, Emergent Parallelism, and Algorithm Research | stellar-group.org | HPX: A cure for performance impaired parallel applications | github.com/STEllAR-GROUP/hpx | This channel is logged: irclog.cct.lsu.edu
Yorlik has quit [Ping timeout: 250 seconds]
K-ballo has quit [Ping timeout: 240 seconds]
K-ballo has joined #ste||ar
<gonidelis[m]> hkaiser: did you happen to check the docs html that i sent you the previous week?
<hkaiser> gonidelis[m]: yeah, let's go with it for now, it's still a RC
<hkaiser> I will do it for the final release
<gonidelis[m]> hkaiser: going over the work stealing scheduler
<gonidelis[m]> is this intended for 1.8?
<hkaiser> gonidelis[m]: no
<gonidelis[m]> ok
<hkaiser> I don't think this will be ready soon
<gonidelis[m]> ok and about the performance test report. where is the 5%-10% perf increase visible ?
<gonidelis[m]> in the report i mean
<hkaiser> gonidelis[m]: the perf test does not use the new scheduler
<gonidelis[m]> ahh so what is this about ?
<hkaiser> I wanted to set up a perf test CI n rostam first (there is a draft PR for this)
<hkaiser> not sure why this perf result was generated, it was an unrelated change, I believe
<gonidelis[m]> oh
<gonidelis[m]> so where do you see the perf results? on your local machine?
<hkaiser> local measurements
<gonidelis[m]> alright
<hkaiser> gonidelis[m]: btw, for task-bench, I want to try creating a numa-aware fork-join executor, let's see if that helps with ARM perf
<gonidelis[m]> arm perf?
<gonidelis[m]> you mean the results you got on fugaku?
<hkaiser> and ookami
<hkaiser> yes
<hkaiser> Nan has very bad perf results on ookami
<gonidelis[m]> ...
<gonidelis[m]> iis the whole prefetching fuzz i 've been hearing around about that numa awareness?
<hkaiser> no, that's different and unrelated
<gonidelis[m]> ok let's please go through it on thursday
<gonidelis[m]> starting getting our hands dirty
<hkaiser> ok
<gonidelis[m]> finally, was the latest apex tagged integrated into master?
<gonidelis[m]> tag*
<hkaiser> not yet, it's still a PR
<hkaiser> #5860
<gonidelis[m]> ok i will go through them rn
<hkaiser> we also need #5864 in the release
<gonidelis[m]> see what's missing
<hkaiser> and your release docs PR, that's it
<gonidelis[m]> hm
<gonidelis[m]> ?
<hkaiser> yes
<hkaiser> it has been deprecated in C++20, so #5864 makes sure it's not used anymore
<gonidelis[m]> the macro name is different though ;p
<hkaiser> is it?
<gonidelis[m]> INIT_FLAG
<gonidelis[m]> FLAG_INIT
<hkaiser> ahh, it's just a typo, thanks
<hkaiser> will fix in the PR
<gonidelis[m]> what's its purpose though
<gonidelis[m]> why initialize atomic_flag ?
<hkaiser> read the cppref page
<gonidelis[m]> why need to ^^
<gonidelis[m]> yes that's a question on cppref
<hkaiser> let's talk Thu
diehlpk has joined #ste||ar
K-ballo has quit [Quit: K-ballo]
diehlpk has quit [Quit: Leaving.]
diehlpk has joined #ste||ar
diehlpk has quit [Ping timeout: 240 seconds]
diehlpk has joined #ste||ar
diehlpk has left #ste||ar [#ste||ar]
hkaiser has quit [Quit: Bye!]
Yorlik has joined #ste||ar
hkaiser has joined #ste||ar
K-ballo has joined #ste||ar
K-ballo has quit [Read error: Connection reset by peer]
K-ballo has joined #ste||ar
K-ballo has quit [Read error: Connection reset by peer]
K-ballo has joined #ste||ar
<satacker[m]> In case of class inheriting `tag_fallback_noexcept` how do I make sure that the tag dispatching takes place when user does incorrect return type overload?
<hkaiser> satacker[m]: how do you overload on the return type
<hkaiser> I don't think that's possible
<satacker[m]> hkaiser: I meant `tag_invoke`
<satacker[m]> * Sorry, I meant
<hkaiser> C++ doesn't allow you to overload functions based on the return type
<satacker[m]> hkaiser: yes, i completely used a wrong terminology. Say tag_dispatching using `tag_fallback_noexcept`, when will the user's `tag_invoke` be called?
<satacker[m]> is the impl
<satacker[m]> are the failing tests
<hkaiser> tag_invoke will be tried first (i.e. used if it is valid), tag_fallback_invoke will be used (if it is valid) only if no tag_invoke's are available
<satacker[m]> hkaiser: Thanks, but how do I make sure the tag_invoke return type is valid, not possible?
<satacker[m]> (The tag_invoke which user implements)
<hkaiser> it wil fail compiling if not
<satacker[m]> Okay, thanks, probably I have done some other mistake, because it compiles, only static_assert fails due to other reasons.
<K-ballo> how can it compile if static_assert fails?
<satacker[m]> static_assert as in tests
<satacker[m]> and same is the case of wg21 implementation... (full message at https://libera.ems.host/_matrix/media/r0/download/libera.chat/44c120333b1982b184b1d8a9a88c6c4f443d4f5a)
<satacker[m]> This is the BAL's implementation
<gonidelis[m]> hkaiser: could you please check your email?
<hkaiser> gonidelis[m]: will do
<gonidelis[m]> Thanks!
<hkaiser> gonidelis[m]: see pm, pls
<Yorlik> o/
<hkaiser> \o
<Yorlik> How's HPX doing these days? I've been quite a bit out of the loop. We're still at 1.7.1 and working on connecting Unreal Engine 5 to the server.
<Yorlik> Seems everyone is busy - gotta run. See you another day :)
* Yorlik waves and fades
Yorlik has quit [Quit: Leaving]
<diehlpk_work> hkaiser, Ok, boost does not support cross compilation using the Fujitsu compiler
<hkaiser> diehlpk_work: nod, do we need the Fujitsu compiler?
<diehlpk_work> hkaiser, Do we have a bug in the hello_world_distributed?
<diehlpk_work> If I run it without mpiexec, it shows 0 loc with 48 cores
<hkaiser> diehlpk_work: if you run it without mpiexec it will create one locality
<diehlpk_work> If I run with mpiexec it shows 0 loc with 48 cores as well, but the cores are printed multiple times
<hkaiser> really?
<hkaiser> that doesn't sound right
<hkaiser> what's you mpiexec parameters?
<diehlpk_work> mpiexec paht to app
<hkaiser> no -n?
<diehlpk_work> I set the parmeter using the the scheduler
<hkaiser> what scheduler?
<hkaiser> srun?
<diehlpk_work> pjsub
<hkaiser> not sure if we support that
<hkaiser> diehlpk_work: is there some documentation on what environment variables this batch system creates?
<diehlpk_work> hkaiser, I need to check that
<hkaiser> alps, slurm, and pbs
<hkaiser> everything needs to be added
<hkaiser> everything *else*
<diehlpk_work> Fugaku uses pjsub
<diehlpk_work> To access the documentation, you need to install the root certificate
<hkaiser> I need the environment variables set by it
<hkaiser> this gives me a 404
<diehlpk_work> Ok, it seems they removed this post
<diehlpk_work> hkaiser, check your email
<hkaiser> ok, I'll have a look, thanks
<hkaiser> is this batch system derived from one the others?
<diehlpk_work> hkaiser, I never heard of this tool before
<hkaiser> me neither
<diehlpk_work> #PJM --mpi "shape=1"
<diehlpk_work> #PJM --mpi "max-proc-per-node=48"
<diehlpk_work> Shoudl give me 1 mpi rank with 48 cores
<hkaiser> nod
<diehlpk_work> They provide 1d, 2d, and 3d node allocation
<diehlpk_work> shape could be 1x2
<diehlpk_work> or 1x2x3
<diehlpk_work> Das ist spannend
<hkaiser> I don't have access to that repo
<diehlpk_work> Wenn wir > 32 numa domain haben dann wird HPX schlechter
<diehlpk_work> Nan laedt dich ein
<diehlpk_work> hkaiser, I assume that hpx does not read the env from the new scheduler correct and therefore a single node run is slower as on Ookami using slurm
<hkaiser> yes, I'm surprised anything works at all
K-ballo has quit [Ping timeout: 250 seconds]
K-ballo has joined #ste||ar
diehlpk_work has quit [Remote host closed the connection]