Every interview subject and all but three survey respondents used visualization regularly to validate the results of runs. For those projects, any problems and productivity bottlenecks that affect visualization have as much impact on overall productivity as poor computational performance during a run.
Such bottlenecks can include problems with available visualization and image conversion software and with finding capable resources for post-processing and visualization rendering. As evidence, one interviewee stated a desire for a large head node with a large amount of RAM to handle rendering. This implies that users' productivity could be greatly enhanced by a dedicated visualization node that shared a file system with computation nodes.
In performance tuning, it is common to hear advice to profile before attempting an optimization, because often the true bottleneck is a surprise. In this paper, we have attempted to find where the true bottlenecks in HPC productivity are by investigating general trends in user behavior and by asking users directly.
Our analysis shows that, in general, HPC user needs are heterogeneous with respect to HPC resource usage patterns, requirements, problems experienced and general background knowledge. These factors combined dictate how individual productivity is viewed and clarifies the motivations for focusing on performance. Our research also gave us insight into the tools and techniques currently available and to what extent they have been embraced in the community.
Future research will evaluate the effectiveness of some of the techniques discussed by users. Based on our findings, we will continue to evaluate HPC community feedback to build a general consensus on the problems that affect productivity and where research time and system money can best be spent to allow these systems to live up to their promise and continue to support leading-edge research.
2 Kepner, J. “High Performance Computing Productivity Model Synthesis,” International Journal of High Performance Computing Applications, 18: 2004. pp. 505-516.
3 PMaC web site for productivity study materials. PMaC Publications. www.sdsc.edu/PMaC/HPCS/hpcs_productivity.html . Aug. 2006.
4 SDSC User Services. SDSC DataStar user guide. website. www.sdsc.edu/user_services/datastar/ .
5 Robson, C. “Tactics -the Methods of Data Collection” in Real World Research, Blackwell Publishing, 2002.
6 Cook, C., Pancake, C. “What users need in parallel tool support: Survey results and analysis,” IEEE Computer Society Press, pages 40– 47, May 1994.
7 Simon, H. “Rational choice and the structure of the environment,” Psychological Review, Volume 63: 2002. pp. 129–138.
8 Pancake, C. “Can Users Play an Effective Role in Parallel Tool Research?” in Tools and Environments for Parallel Scientific Computing, SIAM, 1996.