Datenschutzerklärung|Data Privacy
Impressum

09.11.2020
Martin Pagel

Mo. 09.11.2020 16:00 - 16:45 Uhr, Online: "Data-Parallel Query Processing on Non-Uniform Data" (Jens Teubner, TU Dortmund)

Abstract:
Query compilation is an extremely effective measure to reduce the bandwidth demand of modern in-memory database engines. The same bandwidth reduction would be attractive also for GPU-based systems. But the data-parallel execution model of modern GPUs is at odds with existing query compilation mechanisms: divergence effects nullify the benefits of compiled execution.
To remedy the situation, we propose “push-down parallelism” and “lane refill” as two novel mechanisms which—for different causes of divergence—restore full resource efficiency on modern GPUs.
In the talk, I will first illustrate the high overhead of divergence in commodity settings, then sketch our two anti-divergence techniques.
With a set of experiments, I will show how they can restore efficiency with negligible cost even in adverse scenarios.

Short bio:
Jens Teubner is the head of the DBIS group at TU Dortmund University. His research interests lie in scalable data processing on modern hardware architectures. He holds a PhD from TU München and has been working at IBM Research and ETH Zurich before joining TU Dortmund University in 2013.

Log-in information:
If you are interested in attending the online presentation, please contact web@bifold.berlin !