P4L2. Distributed File Systems

Question 1

What is the upload/download model?

Examples?

Pros/Cons?

Answer 1

Client downloads a file from the server, performs updates on it locally, then uploads the file.

Example: FTP, SVN

Pros

+ local reads/writes at client

Cons

• entire file download/updload even for small accesses
• server gives up control

Question 2

What is the true remote file access model?

Answer 2

Every access to remote file goes to the server, nothing done locally

Pros

+ file access centrilized, easy to reason about consistency, multiple clients can’t overwrite a file at the same time

Cons

• every file operation pays a network latency cost (even reading only since client cannot cache file)
• limits server scalability b/c everything has to go through server

Question 3

What is a stateless file server?

Pros/Cons?

Answer 3

Stateless means the sever doesn’t keep any information (e.g., which clients access which files, how many clients there are, etc). Every request has to be self-contained (include everything it needs to do its work like file name, offset, data).

Pros

+ no resources used on the server side to maitain state (CPU/MM)

+ resilient: on failure, just restart

Cons

• cannot support caching and consistency management (we need state to do this)
• every request self-contained => more bits transferred to describe request

Question 4

What is a stateful file server?

Pros/Cons?

Answer 4

A server that keeps state needed to track what is cached/accessed (e.g., who had portions of the file cache, who has written to a file, etc)

Pros

+ can support locking, caching, incremental operations

Cons

• need checkpointing and recovery mechanisms to handle failure
• overheads to maintain state and consistncy => depends on caching mechanism and consistency protocol

Question 5

What is caching state in a DFS?

Answer 5

Clients can locally maintain a portion of state (e.g., file blocks).

Clients can locally perform operations on cached state (e.g., open/read/write)

A coherence mechanism is require to keep the cached portions of files consistent with the server representation.

Question 6

What is “UNIX semantics”?

Answer 6

Every write is visible immediately

Question 7

What is “session semantics”?

Answer 7

• write-back on close(), update on open()
• easy to reason about, but may be insufficient

Question 8

What is “periodic updates”?

Answer 8

• client writes-back periodically => clients have a “lease” on how long they can used the cached data (not exclusive necessarily)
• server invalidates periodically => provides bounds on inconsistency. easier to correct conflicts b/c they are fewer & smaller
• augment with flush()/sync() API

Question 9

What is “immutable files”?

Answer 9

• never modify, new files created

Question 10

What is replication?

Pros/Cons?

Answer 10

Each machine holds all files

Pros

+ load balancing

+ availability

+ fault tolerance

Cons

• writes become more complex (sychnronously write to all, or write to one then propagate to others)
• replicas must be reconsiled (e.g., voting)
• scalability (machines only get so larage)

Question 11

What is partinioning?

Pros/Cons?

Answer 11

Each machine has a subset of files

Pros

+ availabiity vs single server design

+ scalability w/ file system size

+ single file writes are simple

Cons

• on failure, lose portion of data
• load balancing harder; if not balanced, then hot spots possible