|
Conquest-2
is a three-year
NSF-funded collaborative research grant with Florida State
University. In conjunction with Andy Wang of Florida State
University, we are developing a file system that can simultaneously
provide high file systems performance and energy efficiency. The
key for reconciling these generally conflicting goals is based on the
observation that RAM is becoming an inexpensive resource that can
deliver all file system services with the single exception of large
storage capacity. In essence, instead of storing most files on
disks, Conquest-2 stores them only in RAM. Also, it stores all
metadata in RAM. Since most files can then be accessed
directly from RAM, there is no need to keep the disk spinning.
Only large files need to be kept on disk, and such files will be
accessed infrequently. While RAM running in its normal mode
consumes a great deal of power, RAM can be put into an extremely
power-efficient mode and returned to its normal mode much more rapidly
than disk can.
Our research suggests that one can achieve better power savings than LRU caching by instead using RAM with battery backup as the primary persistent storage for most files, and providing two separate data paths to memory and disk storage. By specializing the data paths, the memory data path contains no-disk-related complexity. The disk data path consists only of optimizations for the specialized disk usage pattern. Conquest-2 is based on the original Conquest disk/RAM hybrid file system (NSF grant CCR-0098363). Under the original Conquest, all metadata (e.g., file attributes) and all files smalled than a certain threshold are stored in RAM, while the disk stores only the data content of large files. The performance improvement is achieved by matching the user behavior with physical characteristics, and having two specialized data paths. Conquest-2 will extend the original Conquest to handle multiple disks. Minimizing the energy use for multiple disks involves creating a bimodal distribution of active and idle disks, so we can put idle disks to sleep to conserve power. The idea is to design and implement an energy-efficient RAID that can "gear shift" disks on and off to adapt for changing loads. This energy-efficient RAID also needs to address existing load balancing and failure recovery concerns by exploting user access patterns, applying erasure encoding, and striping data in a non-uniform way. |