00:00 <Yorlik> At least most of the data

00:00 <Yorlik> There will ofc be some cross referencing here and there, but I am trying to avoid it

00:01 <Yorlik> Items get never ever mved, so I can guarantee pointer validity over the objects lifetime

00:01 <Yorlik> That's why I wrote it

00:01 <Yorlik> Speed + pointers valid

00:02 <Yorlik> The OS specific part is only 4 very short functions - still gotta do the Linux part

00:04 <Yorlik> K-ballo: Forgot about the prefetch - that surely comes in too here. I think my numbers are prbably not too far off.

00:06 <Yorlik> Throughput was a bit above 2GB/sec

00:06 <Yorlik> Thats still far below the theoretical max of ~12

00:52 <hkaiser> Yorlik: you can utilize the full memory bandwdth only if you run on more than one core

00:53 <hkaiser> you have to run on many cores, actually

00:53 <Yorlik> That will be fun then later, when its more of a real application, Thanks for the heads-up

00:54 <Yorlik> In the moment I'm just creating little Lego bricks.

10:30 <mdiers_> hkaiser: yt?

11:24 <Yorlik> Does hpx has a query system, like "Give me all id_types of a given type" ?

12:00 <heller> yes

12:15 <hkaiser> heller: does it now?

12:24 <mdiers_> hkaiser: after your commit "Cleaned up code" about a month ago my clang code analyzer runs into an undefined behavior. i have localized the corresponding location and adapted it for me for testing. it is in hpx/traits/brace_initializable_traits.hpp the is_paren_constructible() . i have adapted it like the overlying is_brace_constructible() with a decltype. then it works again.

12:25 <hkaiser> mdiers_: please create a PR, if you would

12:25 <K-ballo> I'm interested in seeing that diff

12:26 <K-ballo> and also curious that we have a list-init trait

12:27 <mdiers_> All right, I'll put it on.

12:27 <hkaiser> K-ballo: that came in with the c++17 serialization

14:01 <aserio> jbjnr_, simbergm, heller: will you be joining the meeting now?

14:01 <simbergm> aserio: yes, thanks for the heads up

14:01 <jbjnr_> trying now

14:01 <simbergm> I would've joined in an hour...

14:02 <aserio> https://lsucct.webex.com/lsucct/e.php?MTID=m893112686fe4e9b7e0f52dfd754ec5c3

14:54 <diehlpk_work> jbjnr_, yet?

15:04 <heller> Grr forgot the time difference as well :(

15:04 <diehlpk_work> slurm-12475133.out:<jemalloc>: Error in munmap(): Cannot allocate memory

15:04 <diehlpk_work> Has anyone seen this before?

15:09 <heller> Oom

15:10 <heller> Or some strange interaction with memory pinning from the networking layer

15:10 <heller> Which parcelport are you using?

15:34 <diehlpk_work> heller, MPI

17:39 <heller> diehlpk_work: yeah, could be the registration. At which state in the program does this happen?

17:40 <diehlpk_work> heller, During loading the restart file

17:41 <heller> could it be a legit OOM error after all?

17:45 <diehlpk_work> No, the node has 64GB and I load a 1GB file

17:46 <zao> On my site we had memory leaks in something below OpenMPI recently, UCX I think.

17:46 <zao> In some type cache or something.

17:46 <diehlpk_work> I will give it a try with tcmalloc

18:33 <heller> diehlpk_work: you should also try with the system allocator