Challenges and Opportunities of Visual Performance Data Analytics

Speaker: Bernd Mohr (JSC)
Date: Friday, 4 December 2015, 13:30-15:00
Session: Performance Tools I
Talk type: Short talk (15 min)

Abstract: Large-scale scientific computing has become a driver of discovery in many scientific fields. However, efficiently utilising today's complex parallel computing platforms is non-trivial. Application developers need to optimise their codes in a continuous laborious process, diverting a significant amount of manpower from the original scientific task. Also, leaving hardware potential unexploited would take a substantial toll in terms of machine time, energy, and ultimately scientific throughput. Although the optimisation is usually aided by performance-analysis tools that collect performance data and analyse it to some degree, the sheer amount of data produced on current large-scale systems presents a serious challenge. Analysing it requires a degree of expertise and intuition that only few scientists possess. Moreover, the rapidly increasing complexity and size of machines are expanding the scale of the problem even further. The proposal is to adapt and apply methods from the emerging field of visual analytics to extract knowledge from large-scale performance data more effiectively. Motivated by the need to find relevant information in huge data sets, visual analytics combines visual data exploration with automated data analysis. This will make it easier to identify and realise optimisation potential in HPC applications.This requires to bring together experts from the fields of parallel performance analysis and visual analytics.