We’re less than two weeks away from the start of the SC13 Student Cluster Competition in Denver. Right now, the 12 student teams are turning over every rock they can find in their hardware and software stack in an effort to eke (and geek) out every last bit of performance.
Time is short. In a week or less, they’ll be packing up in preparation for the trip to Denver and their meeting with cluster competition destiny.
In an effort to give the students (and anyone else interested in optimizing hybrid CPU-GPU systems) an edge, I recently recorded a webcast with Mark Ebersole, NVIDIA’s CUDA Educator/Developer and evangelist.
In the webcast, Mark and I talk a little bit about the history of GPUs in the cluster competition and how Mark (and NVIDIA) became involved in the cluster battles. The conversation then moves onto how to best configure hybrid systems.
We spend quite a bit of time talking about how to optimize code for GPUs and how to figure out if you’re getting all you should be getting out of them.
There are some important tips in this webcast, like how to use profilers to find opportunities for optimization. We also talk about the specific applications in this year’s Student Cluster Competition, with Mark offering up suggestions for packages and tools.
Along the way, we pondered the question of whether high quality aluminum foil is a good substitute for ECC (it’s not, according to Mark) and how students should approach collecting swag on the show floor (get your bag first, and then get the goodies).