Replication in the Harp File System
--- Liskov et al., 1991.

Summary by Petros Maniatis.

- Short summary: Harp adds availability to file systems by using a
  replicated set of servers, consisting of primary servers, backup
  servers and witnesses.  Compared to NFS, Harp achieves better
  performance, even though it adds extra communication.  The trick
  is to keep a log of the operations in main memory and supply each
  server with a UPS.  Therefore, synchronous disk accesses are 
  avoided in Harp.

Basic idea is replicating storage servers through a primary storage
manager. Furthermore, writes are aggregated through the use of the
write-behind strategy, in conjunction with a UPS, to make sure that
power failures don't kill the volatile write log.

Harp is not a file system. Rather, it's a replication extension
written using the VFS interface. It strives to maintain Unix and NFS file
system semantics.

Every file is assigned a server group. A server group consists of a
primary server, a bunch of backups and a bunch of "witnesses". When any of the
primary or backups is disconnected, or partitioned away, a new view of
the server group is created, where one of the witnesses is promoted to
serve as backup. Witnesses only store event logs (i.e., sequences of
modifications on the file system). Logs are played back during recovery to
resynchronize the state of the file system.

Modifications on the file system, as stored in event logs, are applied
by an applier kernel thread. This thread goes through the event log and
calls the appropriate local filesystem calls to have them propagated to disk.