[ Home | Class Roster | Syllabus | Status | Glossary | Search | Course Notes]
serveless management
software RAID
stripe of data partitioned into N-1 data blocks and a parity block - each on a different disk
high bandwidth
reconstruct loss from parity block (Ex-OR of others)
problem with small writes
cooperative caching
assumes fast connections and trusts relationships
potential to outperform NFS
Log structured Files (LFS)
commits file buffer in large chunks
append only file - latest at end of log
inode information is also written at end of log.
complicates read to find log buffer files blocks
imap contains the current log pointers of mobile inodes
can reconstruct imap from log pointers and "deltas"
checkpoints stored in log
uses versions - need to garbage collect
ZEBRA combines RAID and LFS
stores deltas in log file
anything anywhere, migrate towards operation
unix like i-node tree structure.
stripe groups
current and obsolete groups
see figure 3 for cooperative caching
notice use of other client's cache to avoid need to access server
only manager keep index nodes
cache consistency is on a per block basis
when need to write, lock the block - invalidate other cached copies
subsequent reads/writes by other clients will remove lock
uses first write policy to co-locate manager