OBFS: A File System for Objectbased Storage Device

1
OBFS A File System for Object-based Storage
Device
2
Traditional File System

• Traditional file system have been optimized for
  storing files rather than objects. At the OSD
  level, objects typically have no logical
  relationship, presenting a flat name space.
• Manage large amounts of metadata and directory
  information.
• Most systems do not store data contiguously with
  metadata.
• Trade off the reliability for better performance,
  EXT2 caches metadata updates in memory,EXT3 uses
  lazy log write policy.

3
Design Principle

• Enables load balancing and high performance by
  striping data from a single file across multiple
  OSDs. Objects are limited by the stripe unit size
  of the system.
• Asynchronous writes will therefore be cached by
  the client, almost all of the writes to the OSDs
  will be synchronous.
• A flat name space is used to manage the objects
  on each OSD, efficient searching in this flat
  name space is a primary requirement.

4
Design and Implementation

• Flat name space
• No directory and locality information are
  available.
• General purpose file system is inefficient in
  manage very large directories due to do linear
  search.
• Solution
• Use hash tables to organize the objects.

5
Design and Implementation

• Stripe unit size Block size
• Stripe unit size ? Throughput ?Stripe unit size
  ? Parallelism ?
• Block size ?Block size ? Waste storage space for
  small obj
• Solution
• Employs multiple block sizes and use regions to
  keep blocks of the same size together.

6
Regions and Variable-Size Blocks
7
Object Metadata
8
Object Lookup

• Given an object ID, use Object Lookup Table hash
  table mapping to Onode ID.
• Search Region Head List using the upper 16 bits
  of the Onode ID.
• If the onode belongs to a large block region,
  the object data address can be directly
  calculated.

9
Disk Layout Policy

• Decides incoming request to use what type of
  block.
• For large block, find the nearest large block
  region that contains a free block, mark it used
  and write it to that block.
• For small block, FFS-like allocation policy is
  employed.
• If no region has the desired number and type of
  free blocks, the nearest free region will be
  initialized and put into the active region list.