Steve Kincaid

1
XFS eXtended File System

• Steve Kincaid UWB CSS 434

2
Overview

• Developed UC Berkley in 1996 (1)
• Started by Silicon Graphics in 1993 (3)
• 2001/2002 first distributed with Linux
• Built into most versions of Linux today (e.g.
  Ubuntu)

3
Overview (cont)

• Serverless 64 bit network file system
• Supports file system up to 8 binary exabytes
• 1 EB 1,000,000,000,000,000,000 B 1018 bytes
  or 1 billion gigabytes
• Only 16 binary terabytes on 32 bit Linux systems.
• Takes advantage of fast switched LANs
• Improved performance for
• High loads (many concurrent client requests)
• Large amounts of data
• Big files
• Many files

4
Implementation

• Software RAID storage system striping across
  multiple drives / machines provides failsafe
  against HW failure and it provides large scaling
  capability.
• File management is de-centralized. The XFS
  service manages replication of a Manager Map on
  all machines in the XFS pool.
• File identifiers include a field that acts as an
  index into the manager map, and each entry in the
  map identifies the computer that is currently
  responsible for managing the corresponding file.

5
Journaling

• Journaling file system uses a log (journal) to
  keep track of intended changes to the file
  system.
• Fault tolerant to power failure and system
  failure.
• Insures atomicity
• During system recovery, system follows the
  Journal, and system is recovered. System doesnt
  know its recovering.

6
Allocation Groups

• Equally sized linear regions within the file
  system
• File size can exceed this AG (exception to the
  norm)
• Each AG manages inodes and free space separately
• This allows for one file system to span multiple
  machines
• Multiple machines can support parallel IO requests

7
Extents

• Extent is a contiguous area of storage.
• 1 blocks of disc space
• Block size can vary between systems
• Two B trees index the FS extents / allocation
  group
• Indexed length of free extents
• Indexed starting block of the free extents
• Highly efficient location of free space

8
Delayed Allocation of space

• Lazy write algorithm
• Data sits in the data buffer until certain size
  limit, time limit, or file quantity threshold
  (write barrier) is met. Then data is flushed to
  disc.
• Allocated Extent size match since entire file is
  in buffer
• Multiple file writes land in contiguous memory
• Reduce read, write times
• Written together, likely read together
• Reduce fragmentation
• Avoid writing short term files
• Reader / Writer locking of files

9
TMI

• Sparse File support means no space is allocated
  for empty portions of a file.
• Extended attributes multiple name / value pairs
  per file
• Support for ACL
• Separate user / kernel namespaces
• DMA for even higher performance
• Application must manage locking and concurrency!
• Crash recovery is fast.
• Using journal, not restore

10
Pros and Cons

• File size
• File quantity
• RW throughput
• Supports parallelization
• Fault tolerance
• Performance for a large set of concurrent users.

• An XFS file system cannot be shrunk
• Creation and deletion of directory entries can be
  a much slower metadata operation than other file
  systems.

11
Architecture

• Directory Mgr file system namespace
• Space Manager
• Inode and file allocation
• Manage free space
• Transaction Manager
• Write ahead journaling
• Fault recovery
• Atomic data processing
• Volume Manager
• SW RAID controller

(Sweeney)
12
Performance and Scaling
CPU bound Directory blocks exceeded system
allocated cache. PF
13
File Semantics

• RW locking
• Lazy write
• Write ahead journaling
• One copy update
• Session Semantics with File Lock

14
References

• Coulouris, George, Jean Dolllimore and Tim
  Kindberg. Distributed Systems - Concepts and
  Designs. Essex, Eng Pearson Eduction Limited,
  2005.
• Sweeney, Adam, et al. "Scalability in the XFS
  File System." Proceedings of the 1996 USENIX
  Annual Technical Conference. San Diego USENIX,
  1996. 1-14.
• Wikipedia. XFS. 23 Feb 2010. 23 Feb 2010
  lthttp//en.wikipedia.org/wiki/XFSgt.