Current Theses





Distributed GPGPU on Cloud GPU Clusters Martin Schuchardt Thomas Fahringer details


Title Distributed GPGPU on Cloud GPU Clusters
Language Englisch
Supervisors Thomas Fahringer
Student Martin Schuchardt
Description Cloud instances newly offer GPU instances.
Using GPGPU for problems, which can be solved via massive parallel algorithms, may lead to performance gains on appropriate hardware.
The high number of instances rentable from a cloud provide an interesting basis for powerful distributed systems.
Combining both technologies by distributing chunks of a problem to many instances, and using the GPU power of each instance to compute,
could provide an immense computation power if the problem scales well.
  • Becoming familiar with the cloud infrastructure and GPU instances
  • Write some benchmarks for GPUs
  • Perform benchmarks and compare the cloud instance with stand-alone hardware
  • Write code to simplify and automate creation and configuration of cloud instances
  • Write library/broker to distribute a computation over n instances with m GPUs
  • Evaluate and benchmark using an existing deep learning algorithm provided by IIS/LFU
  • Evaluate and benchmark using another algorithm provided by JAIST, optimize and verify with results from the deep learning algorithm
Theoretical skills
  • Interests in parallelizable algorithms
  • Distributed Systems
  • Parallel Programming (OpenCL)
Practical skills
  • Good C and OpenCL knowledge