SC Conference - Activity Details

Paravirtualization Performance and Programming Support for Next Generation HPC Systems

Lamia Youseff  (University of California, Santa Barbara)
Doctoral Research Showcase Session
Thursday,  11:30AM - 11:45AM
Room 17A/17B
Our research has revealed that paravirtualization imposes minimal performance on HPC workloads, while exposing numerous benefits for this field. In this extended abstract, we overview the work we have done in investigating the performance ramifications of deploying paravirtualized system in HPC settings. For that, we used rigorous statistical analysis techniques to detect performance dissimilarity between our paravirtualized system deployed at Lawrence Livermore National Lab (LLNL), and an out-of-the box Linux Enterprise Red Hat deployment. Our results have shown no statistically significant performance degradation in the 95% confidence level. Furthermore, we outline our results in characterizing paravirtualized memory hierarchy, performance boundary, and paravirtualization impact on autotuning linear algebra software systems. The current direction of our research is implementing VSHMEM: a programming-language support to enable efficient memory sharing, synchronization and communication between applications running in distinct virtual machines to enable effective communication between threads in HPC applications running in virtualized environments.
   IEEE Computer Society  /  ACM     2 0   Y E A R S   -   U N L E A S H I N G   T H E   P O W E R   O F   H P C