Distributed File Systems

n Distributed File Systems provide transparent access to files stored on a

remote disk
n Themes:

Distributed File Systems

n Failures: what happens when server crashes, but client doesn’t? Or vice
versa?
n Performance è caching; use caching at both the clients and the server to
improve performance
n Cache coherence: how do we make sure each client sees most up to date
copy?
Arvind Krishnamurthy n Other issues: naming, scalability, security
Examples
Spring 2001 n

n NFS: Sun’s Network File System

n AFS: Andrew File System (CMU)
n Sprite: Berkeley research project; stressed caching
n Coda: CMU research project; stressed availability

DFS properties Simple Approach

n Network transparency n No caching:
n Looks like local file system n Use remote procedure calls to forward every file system request to remote
n Location transparency server
n No idea where the file is located n Example operations: open, seek, read, write, close
n Location independence n Server implements each operation as it would for a local request and
n Name of file shouldn’t change
sends back result to client
n User mobility n Advantage: server provides consistent view of file system to all clients
n Access anywhere n Problems: performance
n Fault tolerance n Going over network is much slower than going to local memory
n Lots of network traffic
n Access all the time
Server can be a bottleneck – what if lots of clients?
Scalability
n
n
Access from many machines
n
Cache
n File mobility
n Can move files from one location to another in a running system
Client S

Network File System (NFS) NFS structure (cont’d)

n History n An NFS file system can be mounted into the namespace of a Unix
machine
n Introduced by Sun Microsystems in 1985
n Sun published the protocol and licensed reference implementation Server exports file system: /usr/local (in /etc/exports)
n Since then, NFS has been supported by every Unix variants and PC- Clients: mount –t nfs nfsserver:/usr/local /usr/local
NFS
n Daemon processes
On the server: a set of nfsd daemons listen/respond to client requests;
n Design goals n
mountd handles mount requests
n Machine and OS independence, no recompilation of applications
n Crash recovery n On the client: a set of biod daemons handles asynchronous I/O for blocks
of NFS files
n Transparent access
n Unix semantics maintained on client
n Network lock manager & status monitor
n Reasonable performance (comparable to local FS) n Providing facilities for locking files over a network

1
 Caching Motivation, part 1: Failures
n Idea: use caching to reduce network load n What if server crashes? Can client wait until server comes
n Cache file blocks, file headers, etc., at both clients and servers: back up, and continue as before?
n Client memory n Any data in server memory, but not yet on disk can be lost
Server memory
n
n Shared state across RPCs:
n Advantage: if open/read/write/close can be done locally, no network n Example: open, seek, read
traffic n What if server crashes after seek?
n Issues: failures and cache consistency n Client does “read”, it will fail

Cache Client n Message retries:

n Suppose server crashes after it does “rm foo”, but before
Cache acknowledging
n Message system will retry – send it again

S
n What if client crashes?
Cache Client n Might lose modified data in client cache

NFS Protocol (part 1): stateless NFS Protocol (contd.)

n Write-through caching: n Operations are “idempotent”:
n When a file is closed, all modified blocks are sent immediately to n All requests are ok to repeat
the server disk. To the client, “close” does not return until all bytes n If server crashes between disk I/O and message send, client can
are stored on disk resend message; server just does operation all over again
n Read and write file blocks are easy
n Stateless protocol:
n “remove file” is difficult: NFS just ignores this and returns a file
n server keeps no state about client, except as hints to help improve not found
performance
n Failures are transparent to client system
n Each read request gives enough information to do entire operation:
n Suppose you are an application, in the middle of reading a file,
n ReadAt(inumber, position), not Read(openfile) server crashes. Options:
n When server crashes and restarts, can start again processing n Hang until server comes back up

requests immediately, as if nothing happened n Return an error? Problem is most applications don’t know they

are talking over the network

n NFS uses both options: can select which one

Cache Consistency Sequential Ordering Constraints

n What if multiple clients are sharing the same files? n Cache coherence: what should happen?
n Easy if they are reading: each agent gets a copy of the file n What if one CPU changes file, and before it’s done, another CPU
reads file?
n What if one is writing? How do updates happen?
n Every operation takes time; actual read could occur anytime
n NFS has write through policy: between when system call is started and when it returns
n If one client modifies file, writes through to server n Assume what we want is distributed to behave exactly the same as
n How does other client find out about the change? if all processes are running on a single Unix system
n If read finishes before write starts, then get old copy
n NFS: client polls server periodically to check if file has changed
n If read starts after write finishes, then get new copy
n Polls server if data hasn’t be checked in last 3-30 seconds
n Otherwise, get either new or old copy
n When file is changed on one client, server is notified, but other
n Similarly, if write starts before another write finishes:
clients use old version of file until timeout
n May get either old or new version
n What if multiple clients write to same file at the same time?
n NFS: if read starts more than 30 seconds after write finishes, get
n Can get either version (or parts of both). Completely arbitrary. new copy. Otherwise, get partial update

2
 NFS Summary Andrew File System
n Pros: n AFS (CMU, late 80’s) à DCE Distributed File Systems
n Simple n Callbacks: server records who has copy of file
n Highly portable n Write through on close
n If file changes, server is updated on close
Server then immediately tells all those with the old copy
Cons:
n
n
n Session semantics: updates visible only on close
n Sometimes inconsistent
n Unix: updates visible immediately to other programs who have the file
n Does not scale to large number of clients open
n AFS: everyone who has file open sees old version; anyone who opens
file again will see new version
n AFS:
n On open and cache miss: get file from server, set up callback
n On write close: send copy to server; tells all clients with copies to fetch
new version from server on next open

AFS (contd.) Coda and disconnected operation

n Files cached on local disk: n AFS users often go a long time without any communication
n NFS caches only in memory between their desktop client and any AFS server
n What if server crashes? n Coda says: “why can’t we use AFS-like implementation
n Lose all your callback state when disconnected from the network?”
n Reconstruct callback information from clients
n On an airplane
n Pros: n At home
n Relative to NFS, less server load
n During network failure
n Disk as cache à more files can be cached locally
n Callbacks à server not involved if file is read-only n Issues
n Cons: n Which files to get before disconnection
n Fast LANs, local disk is slower than remote memory n Consistency
n Central server is a bottleneck
n Performance (all data + misses goes to server), availability

Hoarding Consistency
n AFS keeps recently used files on local disk n What if two disconnected users write the same file at the
n Most of what you need will be around same time?
n No way to use callback promises since server and client cannot
n Users can specify “hoard lists” to tell Coda to cache a communicate
bunch of other things even if not already stored locally n Coda’s solution: cross your fingers, hope it does not
happen, and pick up pieces if it does
n Log of changes kept while disconnected
n System can also learn over time which files a user tends to
n Apply changes upon reconnect
use
n If conflict detected, try to resolve automatically, else ask the user
n In practice, unfixable conflicts almost never happen