aserio changed the topic of #ste||ar to: 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/
jaafar has quit [Ping timeout: 260 seconds]
hkaiser has quit [Quit: bye]
jaafar has joined #ste||ar
jaafar has quit [Ping timeout: 250 seconds]
K-ballo has quit [Quit: K-ballo]
EverYoung has quit [Ping timeout: 246 seconds]
taeguk has joined #ste||ar
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Ping timeout: 250 seconds]
pradeepisro49 has joined #ste||ar
david_pf_ has joined #ste||ar
david_pf_ has quit [Quit: david_pf_]
david_pf_ has joined #ste||ar
david_pf_ has quit [Client Quit]
pradeepisro49 has quit [Quit: Connection closed for inactivity]
mcopik has joined #ste||ar
mcopik has quit [Ping timeout: 248 seconds]
K-ballo has joined #ste||ar
hkaiser has joined #ste||ar
diehlpk has joined #ste||ar
<heller> doing a test here now: https://gitlab.com/stellar_group/hpx
<K-ballo> 404, no permissions?
<heller> hmm
<heller> it's still importing
<heller> K-ballo: you should be able to see it now
<K-ballo> heller: no, same
<heller> hmmm
<heller> K-ballo: try again
<K-ballo> it asks me to sign-in/up first, which I take as an indication of permission issues
<K-ballo> alright, now it works and is public
<heller> yup
<heller> I have to say, I pretty like the UI
<heller> so far, from brief browsing
<heller> that's pretty cool
<hkaiser> heller: many things are _cool_
<hkaiser> what can gitlab do that's not possible with github?
<hkaiser> also, I don't see yet how changing the repository josting resolves our CI resource problem
<hkaiser> hosting*
<hkaiser> heller: all this jumping onto a new shiny tool without proper discussions and planning is not good, if you ask me
<K-ballo> sounds extreme
<heller> hkaiser: well, the CI system built into gitlab seems to be more suited for our needs
<heller> that's the main motivation
<hkaiser> what CI system is built into gitlab?
<heller> this gitlab runner thing
<hkaiser> and they gibve us unlimited resources?
<heller> no, we provide the resources
<hkaiser> why can't we hook our resources into github?
<hkaiser> in the end it's just a git repo that runner needs
<heller> I haven't found a tool that's capable of doing that with github yet
<heller> and discussion: that's why I wrote the email in the first place
<hkaiser> heller: I think that a) for political reasons, gitlab is inferior to github, and b) it's most confusing to people to have more than one 'main' user-facing repo
<heller> I agree with b)
mcopik has joined #ste||ar
<hkaiser> nobody cares about gitlab, it's a mostly academic thing
<hkaiser> 95% of all OSS happens on github
<hkaiser> in terms of features they provide, github is mostly running after github
<hkaiser> gitlab is running after github*
<heller> from a first look, the service offered by gitlab is superior. If you ask me, there is nothing wrong with having a read only mirror on github, and something else where the development happens
<hkaiser> no
<hkaiser> github has to be th emain read-write repo
<hkaiser> hide your gitlab repo somewhere as read-only for the CI, fine by me
<heller> then this is a dead end.
<hkaiser> what serives are superior? I don't see any
<hkaiser> shrug
<hkaiser> also why is that a dead end if mirroring is seamless as you said
<heller> bryce said mirroring the repo, issues etc. is seamless
<heller> I say, mirroring the code is seamless
<hkaiser> ok, my point still holds
<hkaiser> that's all you need
<hkaiser> I don't care if the tickets are missored properly to a helper repo used for CI purposes only
<hkaiser> mirrored*
<heller> the gitlab UI integrates with the gitlab CI, if we stick with github as our main development, there is no much gain for using gitlab CI, as we loose the integrated reporting facilities (marking PRs as good etc)
<hkaiser> heller: so please tell me - what is so superior
<heller> hkaiser: well, if the CI interface they offer works well for our needs, than this would be one thing
<hkaiser> what are 'our needs' we can't service with github?
<heller> the UI in general offers a few nice features, like resolving conflicts for PRs, creating branches from issues
<hkaiser> all of that can be done with github as well
<heller> hkaiser: using all the resources we have available to test PRs and the master branch
<hkaiser> heller: pls be more concrete - why can't we use 'all the resources we have available'?
<heller> because there is no service that is able to trigger those
<hkaiser> huh?
<hkaiser> don't we do that already?
<heller> where?
<hkaiser> isn't buildbot being triggered ?
<heller> only for master.
<heller> is buildbot able to trigger the machines at CSCS or FAU?
<hkaiser> shrug
<heller> right
<heller> gtg
<hkaiser> show me the docs describing how to do that with gitlab
<heller> [15:49:22] <heller> https://docs.gitlab.com/ee/ci/
<hkaiser> heller: be more concrete, pls - I'm not going to read all of this
<heller> hkaiser: essentially, you have one yaml file describing the build (as we have for circle right now as well), and then we can register runners to execute those, that run on our machines
<hkaiser> I don't see how this can be done without additional support specific to the resources to run on
<hkaiser> so we have to write our own handlers for our resources
<hkaiser> if that is true we could use the same energy/effort to write our own handlers to integrate with github
<diehlpk> What about using Jenkins or something similar for our resources
<diehlpk> I did it before and it was going very well
<K-ballo> I seem to recall two distinct efforts to use Jenkins
<heller> hkaiser: I am not saying it can't be done with github alone, just requires more effort to implement it
<heller> hkaiser: that's what jbjnr and zao try to right now, writing their own handlers
<heller> I'd like to test the gitlab runners nevertheless
<heller> not sure if they'll work when PRs are submited via github though
<heller> from a UI perspective, gitlab and github seem to be more or less equivalent. It's true that more users will look at github first
taeguk has quit [Ping timeout: 260 seconds]
<heller> So yeah, delicate issue.
<heller> If buildbot would only support ssh worker...
<hkaiser> heller: in my book, gitlab looks decent for linux users only, github is much more polished and modern
<heller> Ok ;)
<hkaiser> ;)
<heller> I think it's not worth it to have gitlab just as our CI backend...
<hkaiser> k
<heller> And I'm not sure overall if those gitlab runners are a 100% for us
<heller> +fit
<heller> So looks like a dead end after all
jaafar has joined #ste||ar
<heller> Regarding the gitlab vs. GitHub problematic. There are various big open source projects (llvm, kde, gcc, qt) which only have a read only mirror on GitHub
<K-ballo> llvm and gcc are both still on svn
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Ping timeout: 255 seconds]
<heller> K-ballo: the point I'm trying to make is that the popularity/usage of a oss project does not depend on where the code is hosted or where development is tracked and managed
<K-ballo> I wouldn't expect it to affect popularity/usage, but it does increase the difficulty to contribute
K-ballo has quit [Read error: Connection reset by peer]
<zao> No-one ever got fired for using Github :P
K-ballo has joined #ste||ar
<K-ballo> thunderbird crashed...
<zao> Sounds par for the course.
<K-ballo> where the code is hosted and how development is tracked affects how people get involved with it
<heller> Sure
<K-ballo> kinda how people don't care for submitting issues to boost old track, but don't mind doing so on github
<heller> You can log into gitlab with a github account
<K-ballo> you don't even need to login to boost old track
<zao> Assuming you can tolerate the criminally slow servers and figure out Trac syntax :)
<zao> (my two öre, mirroring repos wherever for testing purposes, super cool)
<heller> I'm not saying we should move away from github just like that. I'm just throwing out a possible solution on how to improve our testing
<heller> If we need to write adapters etc to listen to PRs, we almost wrote our own service anyway
<zao> I'm trying to figure out the state of PRs from a --mirror checkout alone... might have to use GH API for it.
<heller> Yeah
<heller> I mean, triggering builds and update the status for PRs should be relatively straight forward, I think. The hard part is what Jenkins and buildbot does. That is managing resources and matching builds to resources
<heller> The downside of buildbot/Jenkins compared to circle/Travis/appveyor is that once our resources go down, we're stalled
<heller> And none of those come with native slurm support
mcopik has quit [Ping timeout: 255 seconds]
<heller> Could this be something?
<heller> My idea for a solution would be one master that dispatches to the worker via ssh, those just run the scripts in the background and report back. The easiest thing to report back would be the ctest xml output, I guess
<heller> Or push that to cdash, and set the GitHub status with the URL to the cdash result page
mcopik has joined #ste||ar
<hkaiser> heller: as I said, the linux guys believe gitlab is cutting edge
<heller> *shrug*
<heller> hkaiser: I think one argument for gitlab is to not put your information onto servers in the US
<hkaiser> ROFL
<heller> :P
<heller> that's what I hear sometimes, at least
<hkaiser> sure - I couldn't care less, frankly
<heller> me neither
<heller> the only argument for me would have been if our testing could be overall improved
<hkaiser> nod
<heller> and the general workflow
<hkaiser> heller: I'd rather invest some money into circleci
<heller> that'd be an option
<heller> and frankly, I think that's the best shot
<hkaiser> heller: interesting
<heller> doesn't come for free either
<heller> I asked for a quote a few months back
<heller> they'd give us a discount
<hkaiser> how much would that be?
<heller> let me dig it up
<heller> "We can offer a 50% discount. And there is $20 free credits in your account."
diehlpk has quit [Remote host closed the connection]
<heller> that was from october 22nd last year
diehlpk has joined #ste||ar
<heller> that looks like a decent replacement for circle-ci at least, I think. with potentially more bang for the buck. I think circle-ci is rather expensive
diehlpk has quit [Ping timeout: 260 seconds]
jaafar has quit [Ping timeout: 268 seconds]
<github> [hpx] sithhell created fix_action_move_semantics_regression (+1 new commit): https://git.io/vbYVX
<github> hpx/fix_action_move_semantics_regression 9c661f7 Thomas Heller: Fixing local direction function execution and lambda actions perfect forwarding
<heller> hkaiser: so, the local action improvement also broke the action decorations
<hkaiser> heller: uhh, right - didn't think about those
<heller> I am not sure how to fix that :/
<hkaiser> do we have a test for this?
<heller> yes
<heller> one is action_invoke_no_more_than
<hkaiser> ok, will have a look (if I find the time)
<heller> tests.unit.components.distributed.tcp.action_invoke_no_more_than
<heller> the more pressing issue is probably locking hooks
<hkaiser> but that means that direct actions never supported those
<hkaiser> :/
diehlpk has joined #ste||ar
<heller> hkaiser: seems so, yes
<heller> hkaiser: I think I have an easy workaround for now...
<hkaiser> ohh, cool
mcopik_ has joined #ste||ar
<heller> let's see
<heller> hkaiser: so yeah, migration and locking hook are broken
<hkaiser> ok
mcopik_ has quit [Client Quit]
eschnett has quit [Quit: eschnett]
<heller> hkaiser: direct actions did work, they went through the thread_function in all cases
<hkaiser> ok
<hkaiser> I'll have a look - completely forgot about the hooks while applying those changes...
<hkaiser> I think we can change the code to go through thread_function
<heller> a little involved, I think
<heller> and overkill for direct actions
<heller> hkaiser: gentle reminder to look at #3007
<hkaiser> nod
<heller> hkaiser: I would have gone for a solution to bypass the optimizations if an action decoration was specified
<hkaiser> that would work as well - might be even better
<K-ballo> announcement that I'm done touching up the bind_front/back PR
<K-ballo> or not?
<hkaiser> K-ballo: ok, let's merge it asap
<K-ballo> let me double check
<K-ballo> yes, I'm done with it.. I can rebase if necessary
<hkaiser> you decide
<K-ballo> unless you are planning to merge right away, I'd rather rebase
<heller> K-ballo: compile time improvements?
<hkaiser> heller: I'd expect runtime improvements, mainly
<heller> I would have expected no change in runtime
<heller> it's just the compile time logic for the tuple expansion, no?
<K-ballo> heller: no, standard proposal, LEWG had certain opinions on it which I was trying to prove wrong
<K-ballo> I did measure compile time improvements anyhow, there are some but tiny
<hkaiser> heller: the placeholder resolution is not needed at runtime
<heller> K-ballo: can you elaborate on what you tried to prove? a small text in the PR about that would be nice as well
<K-ballo> now I'm sure there are more places where these could be used, since they don't need protection from users passing placeholders, bind expressions, etc
<heller> hkaiser: exactly, and it was handled at compile before as well, in my book
<K-ballo> any place where we used to have to avoid util::bind would be a candidate for bind_front/back
<heller> if the placeholder is in the front or back, right?
<hkaiser> no placeholders
<K-ballo> heller: compelling use cases for bind_back, LEWG drop it because they couldn't think of any
<heller> ahh
<heller> fair enough
<K-ballo> we have a good bunch, but not as many as I was expecting
<heller> bind_front is closer to what other languages refer to as currying, if I am not mistaken
<hkaiser> K-ballo: use cases anyways
<K-ballo> I'll point Tomasz to them once they are merged, but LEWG already shipped the paper forward to LWG by now
<hkaiser> not too late, still
<K-ballo> I somehow missed the first revision of the paper :/ was one of the Konas
<hkaiser> P0443R4 looks like to be ready for implementation now, at least it's closin gin
<hkaiser> K-ballo: even if LWG decides not to have bind_back - us having it is a good thing
<K-ballo> sure, we can keep it
<hkaiser> nod
<K-ballo> we do have the "compelling use cases" after all
<hkaiser> right
<hkaiser> our plan is to be conforming, that doesn't mean we can have more
<hkaiser> can't*
<github> [hpx] K-ballo force-pushed bind_front from d99219a to fdd4271: https://git.io/vFbxT
<github> hpx/bind_front 4404762 Agustin K-ballo Berge: Add implementations of bind_front/back
<github> hpx/bind_front fdd4271 Agustin K-ballo Berge: Replace some uses of bind by bind_front/back
<hkaiser> K-ballo: I'll merge it after tests have cycled
<hkaiser> heller: pls go ahead with merging #3007, I won't have the time to thoroughly investigate things right now
<heller> hkaiser: ok, thanks
<hkaiser> if it breaks something we'll fix it later
<heller> k. shouldn't break anything (tm)
<hkaiser> sure
<heller> or rather: works for me
<heller> hkaiser: what's missing for #2974?
<hkaiser> heller: the comments have not been addressed
<hkaiser> see my last comment
<heller> ok
<heller> i'll ping mikael on monday about it
EverYoung has joined #ste||ar
EverYoun_ has joined #ste||ar
EverYoun_ has quit [Remote host closed the connection]
EverYoun_ has joined #ste||ar
EverYoun_ has quit [Remote host closed the connection]
EverYoun_ has joined #ste||ar
EverYoun_ has quit [Remote host closed the connection]
EverYoun_ has joined #ste||ar
EverYoun_ has quit [Remote host closed the connection]
EverYoun_ has joined #ste||ar
EverYoun_ has quit [Remote host closed the connection]
EverYoung has quit [Ping timeout: 255 seconds]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
<zao> partitioned vector tests take a lot of time to compile, heh.
<zao> 24m5.850s to build for fdd4271, not too bad.
<zao> Scratch that... I accidentally built the wrong commit :D
<K-ballo> 24m to build what? just the tests?
<mcopik> zao: they always been like that
<zao> K-ballo: Everything but examples, I think.
<mcopik> I believe that at point there was a commit introducing explicit declarations of partitioned_vector for int & double to speed up tests
<K-ballo> oh, those are good numbers then
<mcopik> but still, segmented algorithms take some time to build
<zao> Manually driving my containers on the Ryzen box.
diehlpk has quit [Remote host closed the connection]
diehlpk has joined #ste||ar
<hkaiser> yay, they changed , finallygithub to recognize LICENSE_1_0.txt
<K-ballo> hah, I had to modify all my copyright notices
<hkaiser> I opened a ticket with them, took only 6 month or so
<hkaiser> K-ballo: can we delete the inspect-first branch now?
<K-ballo> oh yeah, sorry, will do
<hkaiser> tks
<github> [hpx] K-ballo deleted inspect-first at 5ab60f4: https://git.io/vbYXq
<zao> K-ballo: Build failed.
<zao> Full log, so skim from bottom :)
<K-ballo> zao: is that bind_front branch?
<zao> Should be.
<K-ballo> I ran into similar errors... and fix them
<zao> Commit fdd4271c563aaecb7c1e375d643db6ae30af2e5d
<K-ballo> fix them, rather work around them
<zao> Clang 3.8.1, Debian 9.2, C++14, Boost 1.65.1
<K-ballo> some of the executors are instantiating copy constructors for non-copyable things, but not actually calling them
<K-ballo> I'll check it out
<github> [hpx] hkaiser force-pushed fixing_3027 from 9f24367 to 866c4a7: https://git.io/vbeYH
<github> hpx/fixing_3027 866c4a7 Hartmut Kaiser: Fixing compilation errors
diehlpk has quit [Ping timeout: 248 seconds]
diehlpk has joined #ste||ar
<zao> /tree/hpx/tests/performance/local/serialization_overhead.cpp:35:33: error: no type named 'preprocess' in namespace 'hpx::serialization::detail'
<zao> hpx::serialization::detail::preprocess gather_size;
<zao> hkaiser: ^
<hkaiser> zao: uhh
<hkaiser> what branch is that?
<zao> 866c4a7 you just pushed above.
<hkaiser> circleci is fine
<hkaiser> zao: there were 2 tests still failing, everything else was fine
<zao> Only thing (apart from a test) including hpx/runtime/serialization/detail/preprocess.hpp is src/runtime/parcelset/detail/parcel_await.cpp
<hkaiser> yes
<hkaiser> canyou give me the whole error log, pls?
<K-ballo> transitive includes showing up?
<hkaiser> K-ballo: yah I was fighting this a lot this time
<hkaiser> but as said, circleci works
<K-ballo> ...
<zao> What transitive includes? There's only CPP files in my grep results.
<zao> Unless there's a ghost one defined elsewhere.
<zao> CircleCI hasn't built tests yet.
<hkaiser> zao: ahh, it's a failing test
<hkaiser> will lok
<hkaiser> look
<zao> Great.
<K-ballo> that include used to be in parcel_await.hpp
<K-ballo> bad code was getting it transitively, rather than by direct #include
jaafar has joined #ste||ar
<hkaiser> was just a missing #include
<hkaiser> K-ballo: yah, I tried to resolve this as well
<hkaiser> zao: should be fine now
EverYoung has joined #ste||ar
<K-ballo> zao: the error you see was not an executor instantiating a copy, I changed something that is being called via const to deferred_call
<K-ballo> now if only I could figure which one it is...
EverYoung has quit [Ping timeout: 255 seconds]
<hkaiser> K-ballo: I changed all the then_execute stuff based on jbjnr's attempt in that branch
jaafar has quit [Ping timeout: 260 seconds]
<K-ballo> zao: could you give me a log with -ftemplate-backtrace-limit=0 ?
<zao> Maybe.
<K-ballo> I can't reproduce, I think it's because of return type deduction
<zao> let's see....
<K-ballo> none of the 14 features kick in for msvc (by default at least? I don't know)
<K-ballo> aha, yeah, force defining HPX_HAVE_CXX14_RETURN_TYPE_DEDUCTION and now it reproduces
<K-ballo> so how do I configure HPX for C++17ish ?
<hkaiser> HPX_WITH_CXX17=On
<zao> I'm building everything with -DHPX_WITH_CXX14=On btw, as I asssume Boost is still silly.
<hkaiser> zao: we have fixed that particular issue, I think
<K-ballo> is not a public option?
<hkaiser> K-ballo: it is 'public'
<K-ballo> it doesn't show up when I tell cmake to list options
<hkaiser> not sure what is necessary for cmake to list an option
<K-ballo> option(...) ?
<K-ballo> mmmh:
<K-ballo> Performing Test HPX_WITH_CXX14_CONSTEXPR - Failed
<K-ballo> Performing Test HPX_WITH_CXX14_LAMBDAS - Success
<K-ballo> Performing Test HPX_WITH_CXX14_INTEGER_SEQUENCE - Success
<K-ballo> Performing Test HPX_WITH_CXX14_IS_FINAL - Success
<K-ballo> Performing Test HPX_WITH_CXX14_IS_NULL_POINTER - Success
<K-ballo> Performing Test HPX_WITH_CXX14_RESULT_OF_SFINAE - Success
<K-ballo> Performing Test HPX_WITH_CXX14_VARIABLE_TEMPLATES - Success (cmake feature test)
<K-ballo> Performing Test HPX_WITH_CXX14_DEPRECATED_ATTRIBUTE - Failed
<K-ballo> Performing Test HPX_WITH_CXX14_RETURN_TYPE_DEDUCTION - Failed
<K-ballo> also, we are still testing for HPX_WITH_LIBFUN_EXPERIMENTAL_OPTIONAL?
<hkaiser> K-ballo: yah, looks like we didn't tell cmake about the C++ standards version
<zao> hkaiser: build success of fixing_30x7, now.
<hkaiser> zao: great, thanks
<github> [hpx] hkaiser opened pull request #3039: Consistently use executors to schedule work (master...fixing_3027) https://git.io/vbYDH
<K-ballo> hkaiser: tell me again why the executor customization points use deduced return types?
<K-ballo> the lack of sfinae friendly is poisoning some overload set
<hkaiser> K-ballo: because they have to return whatever the underlying executor returns
<K-ballo> sure, but why do those have to be deduced, as opposed to computed?
<hkaiser> K-ballo: did I screw things up?
<hkaiser> no reason
<K-ballo> there was a reason, something about forward declarations
<hkaiser> don't remember :/
<K-ballo> with return type deductions the execution customization points decay and don't sfinae
<hkaiser> : K-balloahh, I now understand
<hkaiser> K-ballo: I used deduction to avoid sfinae for the cut points
<hkaiser> that made error message slightly better
<hkaiser> K-ballo: pls try to rebase your work onto #3039, that might already fix the issues
<zao> test failures for 3027, forgot to make it verbose so no output :D
<hkaiser> zao: those are unrelated, same happens on master
<zao> ok
diehlpk has quit [Ping timeout: 240 seconds]
<hkaiser> not too bad, however, 6 out of 600 do fail
<K-ballo> same error on top of #3039
<heller> ok, something bad happened
<K-ballo> bind has extra special manual sfinae friendly for one_shot, bind_back/front don't
<hkaiser> K-ballo: do we need to change the CP rt deduction?
<K-ballo> I don't want to go anywhere near those