A large spectrum of data-intensive applications, ranging from small system tools such as CVS and grep, to terascale simulation applications that process huge amounts of scientific data, demand efficient I/O support. These applications run on various computer configurations, from workstations, distributed systems, to large-scale parallel systems. At both extremes the ubiquitous hard disk remains the most cost-effective medium for on-line storage. While the growth of hard-disk capacity nicely matches the rapidly increasing demand for storage, its electromechanical nature is such that performance improvements lag painfully far behind that of processor performance. We continue to observe that the disk bottleneck is worsening in modern computer systems. Thus, improvement in disk I/O performance is critical for improving overall system performance. In the high-end computing arena, where systems are currently rated in teraflops or even petaflops, this rating captures only the performance of the processor and primary memory; I/O performance should also be considered given its significant effect on overall system performance.
In this talk I will present my research on improving disk I/O performance through a better utilization of disk buffer cache. I will describe an integrated caching and prefetching scheme, called DiskSeen, that not only makes access patterns of applications exploitable by the buffer cache, but also makes the data layout of the disk visible and exploitable by the buffer cache. By making disk layout visible to the buffer cache, Diskseen provides functionalities that existing systems do not have. Examples includes random disk accesses being treated differently than sequential accesses so that disk accesses become more sequential, and prefetching being carried out directly on disk blocks using history access information so that metadata and inter-file prefetching is enabled. Using Linux kernel implementations I demonstrate that this technique can significantly improve the performance of a wide variety of applications.
I will also describe my latest research efforts in blendshape animation: reduction of interference among blendshape basis and the direct mapping of facial motion capture data onto blendshape faces. With these problems resolved, the painstaking effort of repeatedly tuning the weights of blendshapes can be significantly saved.
Speaker Bio:
Dr. Song Jiang is a postdoctoral research associate
at Los Alamos National Laboratory. He received his Ph.D
in computer science at the College of William and Mary in
2004, and his BS and MS degrees in computer science from
the University of Science and Technology of China in 1993
and 1996, respectively. His current research is in the
amelioration of the I/O performance bottleneck in
various computer system architectures through improved
memory management and process scheduling. His work on
process/memory scheduling to prevent process thrashing
has been incorporated into the official version of current
Linux kernel.
