Hi Aditya, Thank you for your interest in the machine learning project. As Dr. Kaiser explained, a compiler gathers static information for ML, then ML will select the parameters, such as chunk sizes, for HPX's techniques, such as loop. We have worked on this project since a couple of months ago, and so far we have got interesting results from our implementation. Our focus in the Summer is to implement our technique on a distributed applications. So if you have a background in ML and distributed computing, it would be enough to work on this topic. I am pretty sure that this phase will result in a conference paper as its new and super interesting ;) So if you are interested in this project, go ahead and write your proposal before its deadline.
Best Regards, *Zahra Khatami* | PhD Student Center for Computation & Technology (CCT) School of Electrical Engineering & Computer Science Louisiana State University 2027 Digital Media Center (DMC) Baton Rouge, LA 70803 On Sun, Apr 2, 2017 at 7:04 AM, Hartmut Kaiser <hartmut.kai...@gmail.com> wrote: > Hey Aditya, > > > It would be great if some of you could guide me through the project > > selection phase so that I can make my proposal as soon as possible and > get > > it reviewed too. > > The machine learning project aims at using ML techniques to select runtime > parameters based on information collected at compile time. For instance in > order to decide whether to parallelize a particular loop the compiler looks > at the loop body and extracts certain features, like the number of > operations or the number of conditionals etc. It conveys this information > to the runtime system through generated code. The runtime adds a couple of > dynamic parameters like number of requested iterations and feeds this into > a ML model to decide whether to run the loop in parallel or not. We would > like to support this with a way for the user to be able to automatically > train the ML model on his own code. > > I can't say anything about the Lustre backend, except that Lustre is a > high-performance file system which we would like to be able to directly > talk to from HPX. If you don't know what Lustre is this is not for you. > > All to All communications is a nice project, actually. In HPX we sorely > need to implement a set of global communication patterns like broadcast, > allgather, alltoall etc. All of this is well known (see MPI) except that we > would like to adapt those to the asynchronous nature of HPX. > > HTH > Regards Hartmut > --------------- > http://boost-spirit.com > http://stellar.cct.lsu.edu > > > > > > Regards, > > Aditya > > > > > > > > On Sun, Apr 2, 2017 at 5:21 AM, Aditya <adityasarma...@gmail.com> wrote: > > Hello again, > > > > It would be great if someone shed light on the below listed projects too > > > > 1. Applying Machine Learning Techniques on HPX Parallel Algorithms > > 2. Adding Lustre backend to hpxio > > 3. All to All Communications > > > > I believe I will be suitable for projects 2 and 3 (above). As part of my > > undergrad thesis (mentioned in the earlier email) I worked with Lustre > > briefly (we decided, lustre was an overkill for our scenario as we'd have > > to re organize data among nodes even after the parallel read). I have > > worked with MPI on several projects (my thesis and projects in the > > parallel computing course) and have a basic understanding of all to all > > communications work. > > > > If someone could explain what would be involved in project 1, it'd be > > great. > > > > Also, please let me know what is expected of the student in projects 2 > and > > 3. > > > > Thanks again, > > Aditya > > > > > > >
_______________________________________________ hpx-users mailing list hpx-users@stellar.cct.lsu.edu https://mail.cct.lsu.edu/mailman/listinfo/hpx-users
