ÖÐÎâÄϹËΩһЦ

Actual data is stored in a location pointer (Locator) which is essentially a unique id and allows LOBs to be manipulated without extensive copying.
LOBs are typically stored separately from the data records in whose fields they appear.

Take DB2 Everyplace as an example
    Itis stored in separate LOB table space in an auxiliary table

    LOB Locator (DB2)
        Host variable represents a single LOB value in DB server
        Easily manipulate LOB w/o having to store LOB value to client

Requirements for managing large objects
Object creation/deletion
    lower cost(time) of creating, allocation (deallocation)
Content operation
    whole data/a certain bytes of data operation: seek, read, replace, insert, delete, append
    Good random/sequential access performance
Storage utilization
    Minimize fragmentation
LOB field must be recoverable

Algorithm
    Block-based
        store object in the scattered blocks of disk (slow for sequential read)
    Segment-based
        store object in physically adjacent blocks
        ESM(fixed-size segments)/Starburst(segments of fixed pattern of growth)/EOS(variable-size segments)

1).ESM (EXODUS Storage Manager)
    Support large objects of unlimited size
    All byte operations can be applied
    Data segments are indexed by a B-tree-like structure(Indexed on byte position within the object)

2).Starburst
Extent based allocation
    When size not known
        Allocated for storage double in size until max segment size (Predetermined pattern)
    When size known
        Max size is used to hold
    Last segment trimmed

Efficiency
    Efficient for read, append, byte range replace
    Not Efficient for insertion(deletion) of bytes in middle of object
        Change the size of the long field (entire long field must be copied to new segments)
    suitable for read only object(ex. video, audio)
    
3).EOS
    Object stored in variable-size segments (physically continuous disk pages)
    All of its pages must get filled up except last one
    Positional tree structure
    When byte range delete/insert performed, segments may be broken up    

Disk allocation policy
    binary buddy system:
        manage large fixed-size disk sections of physically adjacent pages
        segments managed as if their size is some integral power of 2

References:
1). Michael Shapiro, Managing Databases with Binary Large Objects, Proc. Of 16th IEEE Mass Storage System Symposium, 1999
2). The Performance of Three Database Storage Structures for Managing Large Objects, SIGMOD 1992