Serverless Network File Systems
Thomas E. Anderson, et al.


XFS is a serverless network file system.  It decentralizes all the tasks
normally assigned to a file server across a cluster of
workstations--actually, the clients themselves.  XFS is appropriate for
clusters of similar computers connected via a high-speed network.
Importantly, the computers must trust each other.

Based on four technologies:
  - RAID:  RAID-style striping of data across workstations
           Two limitations: parity management, RAID is expensive in
               hardware
  - LFS:   LFS addresses problem of small-writes in RAID.
           Overhead associated with log cleaning is the big problem here.
  - ZEBRA: ZEBRA combines LFS and RAID to work well in distributed network
           Issue with Zebra was centralized file manager.  And the fact
           that all segments are striped across all machines in the
           cluster (we can't add a new machine without reducing fragment
           size or increasing segment size).
  - Multiprocessor cache consistency:


Four key maps used to locate data in cluster
  - Manager Map - Globally replicated, maps file index number to a manage
  - imap - split among managers, tracks where file blocks are on disk
  - file directories - maps file name to index number... in on-disk log
  - stripe group map - maps disk log address to set of storage managers


Scalability:  Both XFS and AFS claim to be scalable (and they claim
that the other isn't scalable).  They're using different definitions of
scalable...