Research
Research
My research lies mainly in distributed computing systems with particular interest in data performance, energy conservation, security, and fault tolerance. My postdoctoral research at the Berkeley Lab includes performance optimization for data-intensive scientific applications and energy optimization for datacenter cluster and storage systems, while my doctoral research at the University of Minnesota focused on data access performance in large-scale distributed systems. I also worked for ETRI, a national lab in Korea, as a member of research staff before pursuing my doctoral degree. In ETRI, I participated in various research projects on network security and ATM switching system development. A short summary of my research is as follows:
Energy-proportional Computing
Energy becomes getting more attentions than ever, and the IT community is of no exception regarding this. In 2006, datacenters consumed 1.5% of the total U.S. energy, and it is expected to be 3% of the total in 2011. One major contributor of the severe energy waste in a datacenter is idle power since it is often overprovisioned to handle peak load. Energy proportionality is a new design concept building computer systems that refers to the ability to consume energy in proportion to the given load intensity. At the Berkeley Lab, I worked on developing energy saving algorithms for storage and MapReduce cluster systems to provide energy-proportionality. My research interests in this topic include data layout and replication, load prediction, power optimization in heterogeneous settings, and performance and energy trade-offs.
Related Articles: [SSDBM2011] [EDBT2011] [ICDCS2011] [PDPTA2010]
Network Performance Estimation
In large-scale distribured computing systems, data access can be a critical bottleneck due to node heterogeneity and end-to-end bandwidth scarcity. For efficient data dissemination, network performance estimation is an essential function. Existing estimation techniques are accurate but not very scalable for a large system. At the University of Minnesota, I developed a framework OPEN (Overlay Passive Estimaton of Network performance) that provides scalable network performance estimation, based on sharing of measurements between nodes without toplogical and geographical constraints. OPEN provides a node characterization function to reuse measurements from other nodes, and gossip-based dissemination algorithms for cost-effective measurement dissemination.
The framework has been evaluated with a variety of applications, includingMontage in astronomy and BLAST in bioinformatics. It can be applied to various large-scale systems such as desktop grid systems, peer-to-peer computing systems, and volunteer-based computing systems like BOINC on which various @home projects are running including SETI@home. I am interested in extending the OPEN framework for future cloud systems where multiple clouds exchange a significant amount of data one another, in order to minimize data cost in such a federated clouds environment.
Related Articles: [TPDS2011] [IJIDCS2011][TPDS2009] [ICDCS2008] [CCGrid2007]
Parallel In-situ Indexing
Performance is always of importance in computing systems. Particularly in scientific computing, the size of data can easily go up to tera- or peta-bytes, and as a result, storage I/O becomes a severe bottleneck of performance. While conventional approaches are more related to improving I/O performance, my approach is the side of reducing the frequency of disk accesses for better performance. For scientific applications, the data blocks are often repeatedly accessed for subsequent queries after created, and the data access is dominant to the response time. Since the size of index is much smaller than the data itself in most cases, maintaining index should be beneficial for individual and overall system performance. Extensive measurement studies with a Pixie3D application for MHD (magnetohydrodynamics), have been conducted in NERSC clusters for parallel, in-situ indexing based on FastBit, a bitmap-based indexing technology developed by Berkeley Lab. The implementation has been applied to theADIOS middleware projected by Oak Ridge National Laboratory.
Related Articles:[LDAV2011] [HPCwire(7/21/2011)]
