On Wed, Oct 24, 2018 at 10:14 PM Biddiscombe, John A. <[email protected]> wrote:
> Ahmed > > I'm a little bit frightened by the idea of you working on hwloc+hpx - the > reason is the same as the one that caused problems in gsoc. > Something I have in mind, which would be a good fit with clear direction and specification would be this here: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0796r1.pdf > > * we don't have a well defined plan on what we want to do with hwloc yet. > I would like to clean it up and make it better, but only because it is > missing a couple of functions that I would like - but I don't really know > hwloc that well either. (same as I didn't really know fflib) > > If you are working on a a problem and the tools you have don't quite solve > the problem, then you can tweak the tools to work - for this reason I am > looking at hwloc and asking myself - how do I find out which socket on the > node is 'closest' to the gpu? Should I create a thread pool for GPU > managment on socket 0, or socket 1 (or 2,3.4 ...). I will be working on a > machine with 6 gpu's and 2 processors. it is probably safe to assume that > gpu's 0,1,2 are closest to socket 0, and 3,4,5 to socket 1 - but maybe it's > 0,2,4 and 1,3,5 - I'd like to be able to query hwloc and get this info and > build it into our resource partitioned in such a way that the programmer > can just say "which gpus should I use from this socket" or vice versa. > So I know what I want - but I don't really have a plan yet, and usually, > one knocks up a bit of code, get's something that half works and is good > enough, then tries to make it fit into the existing framework and realize > that something else needs to be tweaked to make it fit nicely with the rest > of HPX. This requires a deal of understanding of the hpx internals and how > it all fits together. > > So in summary, its better for you to find a project that already interests > you (part of your coursework?) and use HPX for that project - this gives > you time to learn how hpx works and use it a bit - then extend it gradually > as your knowledge grows. hwloc/topology is quite a low level part of hpx > and does require a bit of help. If you were here in the office next door to > me, it'd be no problem - cos you could get help any time, but working on > your own is going to be tough. > > I'm not saying do't do it (I did suggest it after all), but it will be > hard to get going. > > I will try to think of some other less low level things we could use help > with. (Ideas Thomas?) > > JB > > ------------------------------ > *From:* [email protected] [ > [email protected]] on behalf of Ahmed Ali [ > [email protected]] > *Sent:* 24 October 2018 14:12 > *To:* [email protected] > *Cc:* [email protected] > *Subject:* Re: [hpx-users] Contributing to HPX > > John, Hartmut, > > Well I can go with that hwloc part. Where should I start? I think I > should start by reading about hwloc from the open-mpi document. Is there > any document about the HPX interface with hwloc? > > Best, > Ahmed Samir > > On Tue, Oct 23, 2018 at 1:46 PM Hartmut Kaiser <[email protected]> > wrote: > >> >> >> John, Ahmed, >> >> >> >> The hwloc integration sounds like a good idea to me. But anything else >> would be fine as well.,, >> >> >> >> Regards Hartmut >> >> --------------- >> >> http://stellar.cct.lsu.edu >> >> https://github.com/STEllAR-GROUP/hpx >> >> >> >> *From:* [email protected] < >> [email protected]> *On Behalf Of *Biddiscombe, John >> A. >> *Sent:* Tuesday, October 23, 2018 3:43 AM >> *To:* [email protected] >> *Subject:* Re: [hpx-users] Contributing to HPX >> >> >> >> Ahmed >> >> >> >> Good to hear from you again. >> >> >> >> >I was a participant in GSoC'18 in HPX but I've failed the program. >> >> >> >> I still feel bad about that. >> >> >> >> >However, I liked the project and the contribution to HPX. I wanted to >> continue contributing to HPX so I want someone to guide me. >> >> Nice. >> >> >> >> >What part should I work on now? and where should I start? >> >> Well, there is the obvious option of continuing the parcelport work, but >> I suspect you want to do something else since we didn’t help you enough >> first time around. I’d certainly like to carry on with that and get it up >> and running. It’s on my list, but I have a full plate already >> unfortunately, so it has to wait if I’m doing it myself. >> >> >> >> There should still be a list of ‘projects’ that were compiled for GSoC >> that you could look through and if something looks interesting, have a go >> at it. >> >> >> >> If you have a project that you are already working on for your studies or >> hobbies - why not try involving that somehow. Experience tells me that if >> someone has a problem to solve and they use a library like HPX to solve it, >> then they get much better results than if they just make up a project and >> try to implement it. The real advantage to this is that you can USE hpx to >> work on a project without having to understand all of hpx under the hood - >> then when you find that a feature you need in hpx doesn’t exist, then you >> atsrt poking around to find a way of implementing it or adding support. >> >> >> >> One thing that is catching my eye at the moment is our interface between >> hwloc (https://www.open-mpi.org/projects/hwloc/) and hpx::topology >> classes. the code in there is not very well maintained and could do with a >> thorough cleaning up. I’d like to be able to create a topology object from >> hwloc and query not just numa domains (which we have implemented in the >> resource partitioner), but also PCI bus connections to GPUs and network >> cards. If I have a 2 socket machine and only 1 gpu - which socket is >> closest to the gpu etc. That hasn’t been implemented and it would be very >> useful. >> >> >> >> Hartmut, Thomas Do you have anything in mind that Ahmed could work on? >> >> >> >> JB >> >> >> >> >> >> >> > _______________________________________________ > hpx-users mailing list > [email protected] > https://mail.cct.lsu.edu/mailman/listinfo/hpx-users >
_______________________________________________ hpx-users mailing list [email protected] https://mail.cct.lsu.edu/mailman/listinfo/hpx-users
