Ficus UCLA replicated File system

1
Ficus (UCLA) replicated File system

• Wide area file system with millions of storage
  nodes
• Stackable layered architecture not that
  pertinent to what are are discussing now
• Non-serializable consistency one-copy
  availability. Updates are applied to a single
  copy and then propagated in the background
• Highly available
• Duplicate directory names possible
• Update versions monotonically increasing
• Is it eventual consistency?
• Multimaster
• State transfer for files, operation transfer for
  directories
• Conflicting file updates are resolved by user user

2
Ficus How would it behave

• For users home directories at ND. Assume that
  Ficus is built on top of AFS
• What access patterns do we expect (wrt sharing)?
• Use the UC Berkeley file access studies for
  reference
• For Grid clusters such as Condor in CSE
  department
• Desktops provide computation/storage
• For a Internet scale storage (their original
  system goal) of 1 million storage nodes and ND
  like access behavior