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/
parsa has joined #ste||ar
mcopik_ has quit [Ping timeout: 264 seconds]
EverYoung has quit [Ping timeout: 240 seconds]
parsa has quit [Quit: Zzzzzzzzzzzz]
<wash>
I heard the word CUDA
parsa has joined #ste||ar
hkaiser has quit [Quit: bye]
parsa has quit [Quit: Zzzzzzzzzzzz]
bikineev has quit [Remote host closed the connection]
<heller>
aserio: who is the organizing genius now?
parsa has joined #ste||ar
<heller>
aserio: also, timezones and shit
<heller>
aserio: how do you conclude that the talk is in one hour though?
<heller>
aserio: you are making me nervous now
akheir has joined #ste||ar
<aserio>
heller: because you just finished it :p
<heller>
ahh, lol
<heller>
yeah, that makes sense
<heller>
I am growing older you know ;)
<wash[m]>
Don't believe I am muted aserio
<heller>
aserio: the mardis gras talk was very special there, where my results graph was rendered when I walked up to the stage
<wash[m]>
Aserio: is my audio working now?
<aserio>
If you are talking we cannot hear it
hkaiser has joined #ste||ar
<aserio>
Give me a second
EverYoung has joined #ste||ar
EverYoung has quit [Remote host closed the connection]
<wash[m]>
Yah I think the issue on your end, hm
<aserio>
We are working on it
EverYoung has joined #ste||ar
<aserio>
wash[m]: say something
vamatya has joined #ste||ar
<zbyerly_>
aserio, can you call me?
akheir has quit [Remote host closed the connection]
hkaiser has quit [Quit: bye]
parsa has quit [Quit: Zzzzzzzzzzzz]
parsa has joined #ste||ar
aserio has quit [Ping timeout: 246 seconds]
aserio has joined #ste||ar
bikineev has quit [Ping timeout: 240 seconds]
akheir has joined #ste||ar
hkaiser has joined #ste||ar
Vir has quit [Ping timeout: 240 seconds]
bikineev has joined #ste||ar
<heller>
hkaiser: just sent a new PDF, could you please try again?
akheir has quit [Remote host closed the connection]
<github>
[hpx] hkaiser created process_error_reporting (+1 new commit): https://git.io/v5hg4
<github>
hpx/process_error_reporting 0f4b982 Hartmut Kaiser: Improve error reporting for process component on POSIX systems
<hkaiser>
heller: still doesn't work :/
<heller>
hkaiser: hmmm
<heller>
you are using acrobat reader, I assume?
<heller>
might be the missing movie files?
<github>
[hpx] hkaiser opened pull request #2910: Improve error reporting for process component on POSIX systems (master...process_error_reporting) https://git.io/v5hgr
<heller>
stupid latex
<github>
[hpx] hkaiser force-pushed process_error_reporting from 0f4b982 to 8698202: https://git.io/v5hg5
<github>
hpx/process_error_reporting 8698202 Hartmut Kaiser: Improve error reporting for process component on POSIX systems
<diehlpk_work>
When dilitation is finished force_int and one other method can be executed and I like to get a future of the two dependent methods for synchronization
StefanLSU has joined #ste||ar
Vir has joined #ste||ar
<heller>
diehlpk_work: par(task)
EverYoun_ has joined #ste||ar
EverYoung has quit [Ping timeout: 255 seconds]
aserio has quit [Ping timeout: 246 seconds]
akheir has quit [Remote host closed the connection]
<heller>
hkaiser: woohoo! look at that massive green!
<K-ballo>
heller: ping
<heller>
K-ballo: pong
<K-ballo>
heller: do you remember the kind of issues you were seeing with cuda and the pack traversal stuff?
<heller>
yes
<heller>
they are fine in general for CUDA
aserio has joined #ste||ar
<heller>
the problem occurs once they are instantiated from a static, file scope context
<heller>
this reminds me ... I need to push this PR
<github>
hpx/cuda_clang 27549ee Thomas Heller: Merge branch 'master' into cuda_clang
<github>
[hpx] sithhell opened pull request #2911: Fixing CUDA problems (master...cuda_clang) https://git.io/v5hXi
<heller>
K-ballo: meaning, you can go ahead with your fixes
<heller>
it's a problem in the EDG frontend of nvcc
<heller>
the only way around is to not compile this stuff with nvcc
<K-ballo>
I only have problems at the moment, not fixes
<K-ballo>
there's some subtle issues with unwrapping when interacting with dataflow, some instantiation point related issues, recursive instantiation, or similar
<heller>
mhm
<heller>
the unwrapping code is very obscur to me :/
<heller>
didn't have the time to fully grasp it yet
<heller>
regarding recursive instantiations
<heller>
this was supposed the major benefit of the new unwrapping code. might be that some specialization or so is missing so you end up with infinite recursion at some point
<K-ballo>
I suspect something of the sort, but both dataflow and unwrapping are cryptic \to follow
<heller>
yeah...
eschnett has quit [Quit: eschnett]
StefanLSU has quit [Quit: StefanLSU]
pree has quit [Ping timeout: 248 seconds]
<K-ballo>
heller: so is #2829 ok to merge now?
<mbremer>
@heller: Just got to my PC. That's interesting about the idle rates. Is that becuase the thread is being suspended while the network controller copies messages over or is it just not being able to hide send latencies?
<mbremer>
Also does tuning HPX_WITH_ZERO_COPY_SERIALIZATION_THRESHOLD matter?
<hkaiser>
mbremer: it should, but might require some parameter sweeps
<mbremer>
But it is presumably only along that one variable? I don't mind running parameter sweeps. Is it application dependent or more of a function of the network?
pree has joined #ste||ar
eschnett has joined #ste||ar
<hkaiser>
mbremer: it decides at what threshold array's are handled using a zero-copy scheme during serialization
<hkaiser>
this adds some overhead for small arrays but helps for larger ones
<hkaiser>
the current threshold is pulled out of thin air ;)
<mbremer>
@hkaiser: kk, well I'll play around with it some then. Thanks!
EverYoun_ has quit [Remote host closed the connection]
<hkaiser>
mbremer: jbjnr mentioned at some point that the current value is way too small for MPI, even more so for IB
EverYoung has joined #ste||ar
<hkaiser>
mbremer: this will have an effectonly if you send arrays of trivially copyable types (i.e. doubles or so)
<mbremer>
That should help then. I'm only sending vectors around.
<mbremer>
I was think it might be interesting to submit something to tune HPX performance?
pree has quit [Ping timeout: 240 seconds]
<hkaiser>
mbremer: nice idea
<hkaiser>
had not seen this before
pree has joined #ste||ar
<github>
[hpx] aserio pushed 2 new commits to add_checkpoint: https://git.io/v5h77
<github>
hpx/add_checkpoint 4526e9b aserio: Adding checkpoint to doxygen
<github>
hpx/add_checkpoint ccf8f8d aserio: Edit ot hpx.idx
patg[[w]] has joined #ste||ar
pree has quit [Ping timeout: 240 seconds]
<mbremer>
@hkaiser: Yeah it's nice, because they can pay one of their people to work on tuning stuff. I might talk to clint about it today as well. I'm not 100% HPX and stampede2 are set up optimally for one another at the moment
<hkaiser>
good thinking
<zbyerly>
mbremer, hkaiser i just got something back from TACC
<zbyerly>
and i quote: "But I should warn you -- nobody here, not even the avx512 instruction set experts -- have done anything at all with gcc on KNL. "
<zbyerly>
are you guys building with Intel?
<zao>
My site only has GROMACS 2016.3 w/ GCC on the KNLs. Everything else is Intel 2017.early
<mbremer>
zbyerly: Everything we've been doing is with gcc/7.1.0
<zao>
No-one insane enough to desire HPX here yet.
pree has joined #ste||ar
<mbremer>
Can you even build hpx with intel 17?
pree has quit [Quit: AaBbCc]
<K-ballo>
mbremer: no, the intel 17 build is broken at the moment
<heller>
mbremer: regarding idle rates, none of the above. We handle the messages in a background thread. The execution of this is currently considered as"idle"
<mbremer>
That's what I remember hearing (vis-a-vis intel).
<heller>
mbremer: regarding bad performance in distributed, the first thing you might want to do is to change small functions, like getter and setter to direct actions
<mbremer>
@heller: Documentation? Right now I'm using channels and then just using a continuation to set what gets pulled out of the channel.
bikineev has quit [Remote host closed the connection]
bikineev has joined #ste||ar
bikineev has quit [Ping timeout: 248 seconds]
<heller>
mbremer: so you only use channels at the moment?
<heller>
mbremer: no other components?
hkaiser has quit [Quit: bye]
<mbremer>
Yup. It's an unstructured stencil kernel. I have submeshes which are components. But stencils only send messages via the channels
<heller>
can you show some code?
<mbremer>
Sure.
<mbremer>
Let me add you to the repo. And I'll try to send you the salient snippets
<heller>
direct actions are currently not documented :/
<heller>
alright
<heller>
that works
<heller>
are you happy with the single node performance now?
<mbremer>
Yeah I think so
<mbremer>
The grain size still has to be a little large for what I would prefer
<mbremer>
also 2 node performance seems good
<heller>
alrighty
<heller>
what should I build, and how would I run it?
<mbremer>
But at 8 nodes it seems to start degrading. But that might be just effects that aren't showing up
<hkaiser>
cool - so you know whom to talk to wrt a job ;)
<heller>
:P
bikineev has quit [Remote host closed the connection]
bikineev has joined #ste||ar
<github>
[hpx] aserio pushed 1 new commit to add_checkpoint: https://git.io/v5jqd
<github>
hpx/add_checkpoint 7189f7d aserio: Expose iteraors for access to the data in checkpoint...
<rod_t>
hkaiser: I'm not sure what exactly is the problem, I pulled stellargroup/phylanx:dev and checked on my local machine. the setuptools is installed for python 3 but not 2. could `python` command pointing to python2 be the cause of the problem? or the absolute path to python3 is used?