hkaiser 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/
diehlpk has joined #ste||ar
<diehlpk>
hkaiser, The last point for 4096 nodes is added and looks good
<hkaiser>
diehlpk: thanks!
<hkaiser>
diehlpk: are all speedup numbers calculated relative to the same base line?
<diehlpk>
Yes, with respect to the sub grids per second one one node on level 14
<hkaiser>
all relative to the mpi time? or mpi times agaist 1 node and LF against their 1 node?
<diehlpk>
Second one
<hkaiser>
diehlpk: would you think that it might be better to calculate all relative to the same (slower) baseline?
<hkaiser>
diehlpk: is one node MPI time comparable to 1 node LF ?
<hkaiser>
or is there a difference?
<diehlpk>
Good questions. I think having it to their own makes it more comaparable how each of them scales
<hkaiser>
most likely the times are similar
<hkaiser>
right, but it does not say anything about which one is actually faster (if the 1 node times are not the same)
<diehlpk>
No, libfabric is 10 seconds slower
<hkaiser>
interesting
<hkaiser>
but why? 1 node does not need any network?
<diehlpk>
I assume different version of ocotiger and hpx?
<hkaiser>
could be - grrr
<diehlpk>
Also having only one run is not really a good measure
<hkaiser>
ok, let's leave them scale against their own base
<diehlpk>
I assume if we have 10 or 11 it would be closer
<hkaiser>
right
<diehlpk>
I can do the plots with mpi as the baseline tomorrow and we can discuss on Monday what we like to have in the paper
<diehlpk>
have you seen, they will ship the pi cluster on Monday :)
<hkaiser>
the legend on figure 4b is wrong
<hkaiser>
yes, I saw that - cool!
<diehlpk>
But see pm
<diehlpk>
yes, I will fix that
<hkaiser>
thanks!
<diehlpk>
Done
<diehlpk>
I moved the table with the alternatives for VC to the appendix
<diehlpk>
I think it would be better to have the results for daint using mpi and libfabric in one Figure on one page
<hkaiser>
diehlpk: ok, makes sense
<diehlpk>
Same for Cori if we have libfabric results
<hkaiser>
yes, I'd prefer that as well
<hkaiser>
possibly even on the same plot
<diehlpk>
Sure, I will do black mpi and blue libfabric and use the same markers for each level
<diehlpk>
This can be done easilt
<hkaiser>
diehlpk: I'll call it a day now - g'night
hkaiser has quit [Quit: bye]
diehlpk has quit [Remote host closed the connection]
<hkaiser>
jbjnr: I applied some edits to your section, hope you don't mind
<jbjnr>
all good
<hkaiser>
ok, got your email
<hkaiser>
sory for butchering it
nikunj has joined #ste||ar
<hkaiser>
jbjnr: 4.3 reads very well now
<jbjnr>
I pushed a few more changes, make sure you have the last version. hkaiser
<jbjnr>
(abut an hour ago)
<jbjnr>
I can write more if help is needed - but I need to be told what/where as I don't know enough about other stuff.
<jbjnr>
diehlpk_work: PS. I have compiled octotiger against the same hpx that you used in the install dir and will test mpi version ready for tomorrow
<jbjnr>
so a version without any of my parcelport changes
<jbjnr>
which should work fine
<jbjnr>
(famous last words)
<jbjnr>
I wish I knew how you computed the speedups so I could generate those plots directly from my output, then I could do the mpi/lf ones again once they run
<hkaiser>
jbjnr: there is that network perf section completely empty still...
<hkaiser>
jbjnr: section 5.3
<jbjnr>
ok
<jbjnr>
hkaiser: any idea if kevin is running LF on cori? results from that would be a nice addition
<hkaiser>
jbjnr: no idea, you may want to ask him
mbremer has joined #ste||ar
<mbremer>
hkaiser: yt?
<hkaiser>
mbremer: here
<mbremer>
this has got to be something stupid on my end. I've deleted the build directory and rerun cmake, but I'm still getting the same error.
<mbremer>
But I can definitely find where HPX_HAVE_SPINLOCK_POOL_NUM is being defined
<hkaiser>
mbremer: urgh
<hkaiser>
mbremer: cmake generates a defines.hpp file inside the build directory that has all of those PP constants defined
<hkaiser>
could it be that you have one of those installed somewhere?
<hkaiser>
config/defines.hpp that is
<hkaiser>
mbremer: usually under <builddir>hpx/config/defines.hpp
<hkaiser>
it would explain things if you had a stale one of those sitting somewhere the compiler can see it
<mbremer>
Yeah, let me try deleting my install directory as well
<mbremer>
yup that was it. Since I install all dgswem dependencies into one folder, I assume the compiler was finding the stale config in <installdir>hpx/config/defines.hpp
<mbremer>
I'll run the test asap and then add the remaining agas functionality to the unit test
<hkaiser>
mbremer: thanks!
<mbremer>
Do you want me to open a PR or just send you the unit test?
<hkaiser>
mbremer: a PR is less work for me ;-)
<mbremer>
PR it is :)
<hkaiser>
nice!
nikunj has quit [Remote host closed the connection]
nikunj has joined #ste||ar
<hkaiser>
nikunj: hey
<nikunj>
hkaiser, hey
<hkaiser>
nikunj: hey
<hkaiser>
I have an idea what you could work on while visiting
<nikunj>
hkaiser, how're you doing?
<nikunj>
hkaiser, great, what is it?
<hkaiser>
thanks, I'm fine, how about you?
<nikunj>
I'm doing good as well
<hkaiser>
nikunj: we have a small project where we develop resiliency APIs on top of HPX, mostly task related
<hkaiser>
nikunj: I'll add you to a private repo for you to see
<hkaiser>
nikunj: what is your github nick (sorry I forgot)?
<K-ballo>
nikunj: you will be visiting LSU? that's awesome!
<nikunj>
it's NK-Nikunj
<nikunj>
K-ballo, yes! I'll be at LSU from 3rd May all the way to 3rd August
<hkaiser>
nikunj: you should receive an email from github
<nikunj>
hkaiser, it would be great if we could video call sometime. It'll be great if you explain me it's structure
<K-ballo>
nikunj: I did something like that many years ago, you'll have a great time there
<K-ballo>
nikunj: where are you from originally? how do you get along with heat and humidity? :P
<nikunj>
K-ballo, ohh.. it's exciting to work on such a project
<nikunj>
K-ballo, I'm from New Delhi (capital of India). I can do well with 45 degree celsius(113 degree fahrenheit). I'm sure I can get along there xD
<hkaiser>
nikunj: sure, let's plan something for this week
<K-ballo>
ah you'll do fine
<hkaiser>
nikunj: the project is simple enough, just look at the examples
<hkaiser>
the code itself is simple as well, you shouldn't have any issues following
<nikunj>
why does the examples provided look similar to the one provided with HPX itself?
<hkaiser>
because they are all written by us ;-)
<nikunj>
:)
<nikunj>
I have a visa appointment on 10th. So I'll be traveling on 9th and 11th
<nikunj>
we can schedule a video call tomorrow or 10th or 12th
<nikunj>
whichever is comfortable for you
<hkaiser>
sure, tomorrow should be fine
<hkaiser>
or we could talk today, if you liked
<nikunj>
sorry, I have a class early morning tomorrow and it's already 1am
<hkaiser>
ok, let's talk tomorrow then - let's play it by ear
<nikunj>
It won't be a productive one, if we do it today
<nikunj>
hkaiser, sure!
<nikunj>
it's great to finally know about the project I'll be working on. Btw, can I work on phylanx as well?
<nikunj>
It'll be great if I can get some exposure in machine learning as well
<nikunj>
I have not worked on any ml based library yet, so it will be a new experience
<mbremer>
hkaiser: Your PR is working on my end :) I'll get the unit test updated ASAP
<mbremer>
thanks for implementing this!
<hkaiser>
mbremer: sorry it took so long
<hkaiser>
nikunj: sure, the hpxr stuff is not too much of work, but it's paid, so we need to deliver
<hkaiser>
should give you plenty of time to look at phylanx
<nikunj>
hkaiser, I see. I'll deliver for hpxr then. I'll work on phylanx in my free time :)
<nikunj>
I just wanted to work on something related to ml since it's the most debated topic these days
mbremer has quit [Quit: Leaving.]
<hkaiser>
nikunj: sure, we'll find something for you
<hkaiser>
nikunj: and really hpxr is not a lot of work, mostly doin gperf runs and write up the results
<nikunj>
hpxr should be fun
nikunj has quit [Remote host closed the connection]
nikunj has joined #ste||ar
<nikunj>
I'll have to go through gperf tutorials.. I'll go through them before I start with work
<hkaiser>
nikunj: no need for this, what we need there is just some overall timings for one or two applications with vs. without hpxr