Guest Speaker: Karsten Schwan
Talk Abstract: This talk presents the opportunities and challenges presented by future memory technologies like non-volatile RAM (NVM) that offer increased memory capacity as well as fast persistent storage. Prior research has focused either on improving memory scalability by replacing DRAMs with PCM or improving persistent storage by using PCM as nonvolatile heap. In resource constrained end user devices like Smartphones and tablets, however, and likely, also for future exascale machine nodes, it is desirable to leverage PCM for both its capacity and persistence properties. Our research, therefore, is exploring how to obtain these 'dual benefits' of PCM. Specifically, we investigate and evaluate the impact of using PCM for its persistence properties on the performance of applications that are using PCM for capacity. We show that current shared last level cache architectures will cause severe impacts on applications requiring increased memory capacity when there are co-runners using PCM for persistence, via increased cache miss rates experienced by 'capacity' applications. In response, we propose novel methods that e.g., use application page contiguity metrics to reduce such misses. We also investigate other software overheads like those relating to memory allocation allocator, then develop methods to reduce them by redesigning allocator data structures. Current results obtained for end devices are now being extended to also consider server systems and applications.
About the Speaker: Dr. Schwan is a Regents' Professor in the College of Computing at the Georgia Institute of Technology. He also directs the Center for Experimental Research in Computer Systems (CERCS), with co-directors from both GT's College of Computing and the School of Electrical and Computer Engineering. Prof. Schwan's M.S. and Ph.D. degrees are from Carnegie-Mellon University in Pittsburgh, Pennsylvania, where he began his research in high performance computing, addressing operating and programming systems support for the Cm* multiprocessor.
His current research ranges from topics in operating systems, to middleware, focusing on information-intensive distributed applications in the enterprise domain and on data-intensive codes in high performance systems. Technical topics of current interest include (1) high performance I/O for petascale machines, (2) scalable techniques for virtualizing and managing future many-core and datacenter platforms, (3) efficient methods for managing large-scale systems, including new techniques for runtime performance and behavior monitoring and understanding, (4) middleware for high performance data movement and 'in transit' data processing and for QoS-sensitive data streaming in pervasive and wide area systems, and (5) experimentation with representative applications in the HPC, enterprise, and mobile domains.
Host: Rajiv Ramnath