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/
<ct-clmsn>
@wash: ah right, containers!
<ct-clmsn>
@wash: so this is an nvidia container yard?
<zao>
I hear some fools are considering sticking GPUs into my cloud.
<ct-clmsn>
i'm getting compile time errors from master, compiler can't find HWLOC_OBJ_NUMANODE
<zao>
The enumerand should be reasonably old, git blame says late 2014.
<zao>
How old is your hwloc?
<hkaiser>
ct-clmsn: yah, jbjnr added code requiring a new version of nwloc
<hkaiser>
we need to protect that somehow - will talk to John tomorrow
<ct-clmsn>
@hkaiser: ah ok, i remembered something about that happening this week - wasn't sure if ya'll had tested for that
<hkaiser>
ct-clmsn: I forgot to bring it up during review :/
<hkaiser>
sorry for that
<ct-clmsn>
np
<ct-clmsn>
going through the ast transform code
<ct-clmsn>
@hkaiser: i'm going to try to glue my stuff into the mix - it's going to look *really* weird b/c it's probably not going to look very much like the hpx coding style
<jbjnr>
it doesn't get a lot of air-play - as they say ...
<zao>
jbjnr: Say that hypothetically I'd want to shove in results into that there shiny CDash instance. How does one actually do so?
<jbjnr>
ctest Submit ?
<zao>
(I'm going to run to catch the bus now, but still curious in absence)
<jbjnr>
-D or something
<jbjnr>
experimental Update, Build, Submit and the results will appear in the dashboard
<jbjnr>
if you use the hpx project url. Don';t submit other stuff there please :)
<jbjnr>
look in cdash setup or something in the hpx project root somewhere - the urls should be setup already for submission
patg[w] has joined #ste||ar
<jbjnr>
msimberg: I volunteer to do an HPX release once, so aserio sent me that doc, so, I unvolunteered immediately and he did it :)
<jbjnr>
around 0.99999 or thereabouts
<msimberg>
jbjnr: step 29 is at least easy
<jbjnr>
lol
<aserio>
:)
<msimberg>
btw jbjnr: 23 = the number of days since you shaved?
<aserio>
jbjnr: Why the need for a release?
<msimberg>
aserio: I was just asking in general
<jbjnr>
no. 23 was the stack trace depth between calling then_execute and the function actually cally async_execute. Fortunately, hartmut told me I was emulating the wrong async and now it's only a coupl of forwarding calls before the function is dispatched.
<msimberg>
ah, almost...
gedaj has quit [Read error: Connection reset by peer]
<aserio>
msimberg: Yea, we normally do a release before SC... This year we have been kept quite busy with new projects
<jbjnr>
aserio: no need. msimberg was asking how it's done that's all - he comes from a commerical background and loves doing that sort of stuff apparently.
<jbjnr>
I think he wants to volunteer for 1.1
<jbjnr>
:)
<aserio>
I don't think we would stop him...
<msimberg>
mmh, loves
<patg[w]>
Rule 1. Here don't mention anything that you don't want to be volunteered for
<jbjnr>
Rule 2 - don't talk about Rule 1
parsa has joined #ste||ar
<msimberg>
well, I did ask how I could help so I guess I'd be in trouble even without rule 1
<patg[w]>
aserio: did you see the bibs work in the journal format now
<jbjnr>
msimberg: it's about a month now and I cannot wait until halloween is over and I can shave it off. Unfortunately, everyone else seems to like it (family etc).
<msimberg>
jbjnr: is it already the correct shape?
hkaiser has joined #ste||ar
gedaj has joined #ste||ar
<jbjnr>
I've changed plans and will instead do Wolverine from X-Men (probably). But without the muscles
<aserio>
patg[w]: I saw you email... let me try to build it
eschnett has joined #ste||ar
<aserio>
patg[w]: You said you were able to build it?
<patg[w]>
aserio: yes
<patg[w]>
Are you having problems
<aserio>
yea
<patg[w]>
hmm
<patg[w]>
And you pulled the changes
<aserio>
are you using the make file?
<patg[w]>
yes
<aserio>
Ok let me see if that works for me
<aserio>
I am just using the MikTex GUI
<patg[w]>
It seems they should both work
<aserio>
Not necessarily, the make file often is using specific packages to compile things
<aserio>
patg[w]: It seems to have worked...
<patg[w]>
aserio: hopefully we won't need it but I have a feeling we will not be accepted to ASPLOS
<aserio>
Yea, we feel the same
<patg[w]>
aserio: Bryce rebuttal is very good though
<aserio>
It was!
<patg[w]>
aserio: I'll see you at SC right?
<aserio>
Yep!
parsa has quit [Quit: Zzzzzzzzzzzz]
parsa has joined #ste||ar
parsa has quit [Client Quit]
parsa has joined #ste||ar
parsa has quit [Quit: Zzzzzzzzzzzz]
<heller>
msimberg: why should the mask change?
<heller>
The other code is needed if no affinity mask was set
<msimberg>
heller: it may be that I don't know enough about what exactly rp.get_pu_mask returns, so apologies if this is a silly question
<heller>
There are no silly questions
<heller>
I might have gotten it wrong as well
<msimberg>
but at least if you go to the threads::set case, mask will have exactly one bit set in the num_thread position?
<msimberg>
and if you've disabled that pu you'll be stuck in the while loop
<heller>
Did the test succeed if you update the masks in each iteration?
<msimberg>
yes
<heller>
Hmmm
<heller>
Ok, good catch
<heller>
msimberg: excellent catch!
<msimberg>
so I unreverted most of your changes and changed that to update the mask and it's happier
<heller>
That's the missing piece
<heller>
You are of course correct
<msimberg>
the only remaining thing was if you use the local_priority_queue_scheduler it hangs in some cases when going over numa domains
<msimberg>
because only the first pu will steal across domains
<heller>
Yes
<heller>
That's a general problem
<msimberg>
hence I was asking about the rp.get_pu_mask... is it necessary? or would it be enough to do just to set the one bit?
<heller>
Because we rely on stealing here...
<msimberg>
thinking of the case when a lot of pus are disabled and there might(?) be unnecessary work with the rp.get_pu_mask
<heller>
Yes, we need the mask
<msimberg>
so in what cases would the first pu mask be empty?
<heller>
There is this silly thing with logical and physical numberings and all this mess
<msimberg>
hmm, so other way around: when would threads::set be wrong?
<heller>
If the user runs the hpx application with the --hpx:bind=none parameter
<heller>
If the physical id is not the same as num_thread+offset