Why High-Performance Systems Need a Little Bit of LUC
Submitted by hebertem on Wed, 2015-02-04 11:57
Kirk W. Cameron
Colloq: Speaker Institution:
Colloq: Date and Time:
Wed, 2015-04-08 10:00
Building 5700, Room F234
Jeffrey S. Vetter
Colloq: Host Email:
In 1936, Harvard University sociologist Robert Morton wrote a paper entitled "The unanticipated consequences of purposive social action", where he described how government policies often result in both positive and negative unintended consequences. The lesson from Morton's work was that unexpected consequences in complex social systems, at the time relegated to theology or chance, should be evaluated scientifically. Independent groups typically design the components of HPC systems. Hard disks, processors, memories, and boards are eventually combined with BIOSs, file systems, operating systems, communication libraries, and applications. Today's components also adapt automatically to local conditions to improve efficiency. For example, processors and memories can vary their frequencies in response to demand. Disks can vary their rotation speeds. BIOSs and OSs can adapt their scheduling policies for different use cases. Since the performance effects of local hardware and software management are largely unknown, these potentially valuable features are often disabled in high-performance environments. And unfortunately, while we assume that disabling these features will have positive consequences, Morton teaches us that relegating performance behavior to chance is just as likely to result in negative consequences. For example, there is mounting evidence that when processors are fixed at the highest frequency (i.e., disabling dynamic frequency scaling), performance can worsen. In this presentation, I will revisit the conventional wisdom that "faster is always better" for processor speeds in high-performance environments. In essence, through exhaustive experimentation, we can demonstrate quantitatively that slowing down CPU frequency can speed up performance as much as 50% for some I/O intensive applications. For the first time, we have identified the root cause of slowdowns at higher frequencies. I will describe how the LUC runtime system Limits the Unintended Consequences of processor speed in high-performance I/O applications. Our work also motivates the need to reject chance as an explanation of performance and revisit first principals so we can design systems that truly offer the highest performance.
Colloq: Speaker Bio:
Kirk W. Cameron is Professor and Associate Department Head of Computer Science in the College of Engineering at Virginia Tech. The central theme of his research is to improve power and performance efficiency in high performance computing (HPC) systems and applications. More than half a million people in more than 160 countries have used his power management software. In addition to his research, his NSF-funded, 256-node SeeMore kinetic sculpture of Raspberry Pi's was featured at SIGGRAPH 2014 in Vancouver, B.C. and is scheduled for multiple exhibitions in Washington D.C. and New York in 2015.