Thanks Ryan for the pointer, this is awesome work, I am looking forward to this becoming part of the Julia release in Q3.
Sebastian On Thu, Aug 20, 2015 at 3:34 PM, Ryan Cox <ryan_...@byu.edu> wrote: > Sebastian, > > This talk from JuliaCon 2015 discusses progress on OpenMP-like threading: > Kiran Pamnany and Ranjan Anantharaman: Multi-threading Julia: > <http://youtu.be/GvLhseZ4D8M?a>http://youtu.be/GvLhseZ4D8M?a > > Ryan > > > On 08/19/2015 02:42 PM, Sebastian Nowozin wrote: > > > Hi Julio, > > I believe this is a very common type of workload, especially in scientific > computing. > In C++ one can use OpenMP for this type of computation, in Matlab there is > parfor. From the users perspective both "just work". > > In Julia, I have not found an easy and convenient way to do such > computation. > The difficulties I have experienced, trying to do this with distributed > arrays and the Julia parallel operations: > > 1. Having to prepend @parallel before every import/require so that all > parallel processes have all definitions. > 2. Working with the distributed arrays API has given me plenty of > headaches; it is more like programming with local/global contexts in OpenCL. > 3. (I believe this is fixed now.) There were garbage collection issues and > crashes on Windows when using distributed arrays. > > What would be very convenient is a type of OpenMP like parallelism, really > anything that can enable us to write simply > > function compute(X::Vector{Float64}, theta) > n = length(X) > A = zeros(Float64, n) > @smp for i=1:n > A[i] = embarassing_parallel(X[i], theta); > end > A > end > > Where @smp would correspond to "#pragma omp parallel for". > I know this may be difficult to implement for a language as dynamic as > Julia, but it is hard to argue against this simplicity from the users' > point of view. > > As Clang/LLVM now support OpenMP (https://clang-omp.github.io/), one > perhaps can recycle the same OpenMP runtime for such lightweight > parallelism? > > Thanks, > Sebastian > > > On Wednesday, 19 August 2015 19:03:59 UTC+1, Júlio Hoffimann wrote: >> >> Hi Kristoffer, sorry for the delay and thanks for the code. >> >> What I want to do is very simple: I have an expensive loop for i=1:N such >> that each iteration is independent and produces a large array of size M. >> The result of this loop is a matrix of size MxN. I have many CPU cores at >> my disposal and want to distribute this work. >> >> In the past I accomplished that with MPI in Python: >> <https://github.com/juliohm/HUM/blob/master/pyhum/utils.py> >> https://github.com/juliohm/HUM/blob/master/pyhum/utils.py Whenever a >> process in the pool is free it "consumes" an iteration of the loop. What >> exactly the @parallel macro in Julia is doing? How can I modify the code I >> previously posted in this thread to achieve such effect? >> >> -Júlio >> > >
