HPC is much stronger than any virus!
Hard times that we have to live
I don’t think any of us are going to disagree, but we must show it to the world. We have to demonstrate, each of us (from home), that science and HPC are not going to stop due to coronavirus. This Summer of HPC edition will always be remembered because of coronavirus, but we must do an effort to make it also remembered because of all the great work and wonderful findings in HPC.
I’m Pablo and I should be enjoying Paris’s heat by now. However, I’m suffering (enjoying) the Murcia’s (Spain) one. Well…Murcia isn’t a bad city at all, is it?
As you can see, I love photography, but also computer science and HPC. I’m 22 and I graduated from computer science last year and I have recently finished my MSc in computer science, more precisely, in high performance computing (HPC), both at the University of Murcia. The project I will be working on is “Matrix exponentiation on GPU for the deMon2k code“.
What I have been working on
For the past two years, I have been working very intensively in the HPC world. My first work was focused on optimizing Ant Colony Optimization algorithm applied to the TSP problem (see image below). I started with a highly efficient, parallelized and vectorized implementation and, as part of my bachelor thesis, I improved the performance, obtaining a ~300x speedup on Xeon Phi KNL. Before me, running the code could take 12 days, and now it would take 1 hour. Not bad, right? Anyhow, since my bachelor thesis I have been working on this code a lot, so the comparison now would look extremely ridiculous. I also have written a paper about it which I expect to be published soon!
This year I have been working in neural networks. I have been studying the performance of a Caffe (a machine learning framework) implementation which uses PHAST library (which is somehow similar to OpenCL). This Caffe version is coded once, but we can target both CPU and NVIDIA GPUs without touching the code at all (just changing the compilation flags!). However, the performance was poor, so my job was to understand why and fix it. The short story is that the problem was some Caffe layers which were working so bad with PHAST, because the implementation was very poor.
I’m pretty excited about the project I’m going to work on. I always wanted to know how a supercomputer was managed and how is it to have access to one of them. I am excited to start working and to evaluate the performance of the GPUs in the Jean Zay machine. I hope to obtain great performance and exploit this huge computations as much as I can!