Hackathon wrap-up

As mentioned in a previous post, we (me, postdoc Abhishek, and student Grant) were in Edinburgh attending a hackathon run by DiRAC, the computing arm of the UK STFC funding agency, and Codeplay, an Edinburgh-based company with expertise particularly in the Intel architecture of hardware and software.

At the start, we had our code (Sky3d) which runs on CPU and makes extensive use of a CPU-based fast Fourier transform library (fftw).  By the end, we had, with extensive help from the DiRAC and Codeplay engineers, a code which was mainly hosted on the CPU but which offloaded parts of its operation to a GPU coprocessor.  In particular, we were able to interface to a custom set of fft routines which are part of the Intel MKL library and which we were able to use fftw wrappers for.  This was all pretty exciting - we are using the latest hardware, with very new compiler technology, to run our code.  The hope is that it should end up running so much faster that we will be able to attack physics questions that are just too hard / time consuming to answer right now. The down side is that so far, the GPUified code actually runs slower than the CPU-only code.  This is presumably because it is spending time shunting too much data between the CPU memory and the GPU memory and not enough time powering through calculations on the GPU.  So we still have a lot of work to do to get a fully optimised GPU version of Sky3d, but at least we have started now.

At the end of the three day event, Grant and I went to a chip shop.  Grant tried the local delicacy of a deep-fried battered Mars bar.  He liked it.  I didn't fancy one but have tried one before and think they're pretty good! 

Grant with a deep-fried battered Mars bar