Conquest: Better Performance Through A Disk/Persistent-RAM Hybrid File System

1
Conquest Better Performance Through A
Disk/Persistent-RAM Hybrid File System

• A. Wang, Peter Reiher,
• Gerald J. Popek, Geoffrey H. Kuenning
• USENIX Annual Technical Conference, June 2002

2
Introduction (1/2)

• Background
• Over 25 years, long-term storage has been
  dominated by rotating magnetic media
• Today, price of RAM is rapidly dropping
• Becoming affordable to make whole file system
  with RAM
• Consideration
• Rotational delays are not relevant to persistent
  RAM storage
• Existing file system designs are suitable in
  this new environment ?

3
Introduction (2/2)

• Conquest
• Designed to address this environment (no
  rotational delay)
• To smooth the transition from disk-based to
  persistent-RAM-based (battery backed) storage

tape
magnetic disk
RAM
Conquest (disk/RAM hybrid)
4
Conquest FS Design (1/5)

• Media-usage strategy
• Observation from recent studies reveal that
• Most files are small
• Most accesses are to small files
• Most storage is consumed by large files, which
  are often accessed sequentially
• Strategy in Conquest
• All small files, metadata, executables, and
  shared libraries goes to persistent RAM
• Disks hold only the data content of large files

5
Conquest FS Design (2/5)

• Media-usage strategy (contd)
• Choice of small files
• By using simple threshold
• Only the data content of files above the
  threshold (currently 1MB) are stored on disk
• By enlarging this threshold, Conquest can
  incrementally use more RAM storage as the price
  of RAM declines

RAM
RAM
metadata
metadata
metadata
metadata
contents
contents
contents
DISK
DISK
contents
Files lt 1MB
Files gt 1MB
6
Conquest FS Design (3/5)

• Metadata representation
• In core file metadata
• Removes nested indirect blocks from the commonly
  used I-node design
• Doesnt use v-node data structure provided by VFS
• Memory addresses of metadata is used as unique
  IDs.

7
Conquest FS Design (4/5)

• Metadata representation (Contd)
• Directory metadata
• Variant of extensible hashing is used for
  directory representation
• Directory structure is built with a hierarchy of
  hash tables, using file names as keys

8
Conquest FS Design (5/5)

• Metadata representation (Contd)
• Large-file metadata
• To keep track of disk storage locations
• Conquest maintains a dynamically allocated doubly
  linked list of segments
• Although linear search is needed, its simplicity
  and in-core speed outweighs its algorithmic
  inefficiency

9
Conquest Implementation

• Conquest prototype
• Operational as a loadable module (Linux 2.4.2)
• Uses modified VFS
• Omission in in-core data path (RAM)
• Checking the status of buffer cache
• Faulting/prefetching pages from disk
• Flushing dirty pages to disk
• Performing garbage collection
• Omission in disk data path
• On-disk metadata chasing
• Seek-time optimizations for small files

10
Performance (1/4)

• Experimental platform
• 1GHz Xeon processor, 256KB L2 cache,Linux 2.4.2
• Memory 2GB (4X512MB)
• Disk 73.4GB 10,000 RPM, Seagate
• Conquest is compared with..
• ext2
• reiserfs
• XFS (by SGI)
• ramfs (by Transmeta)

11
Performance (2/4)

• Sprite LFS small-file benchmark

12
Performance (3/4)

• Sprite LFS large-file benchmark (1M)

13
Performance (4/4)

• Sprite LFS large-file benchmark (100M)

14
Conclusion

• Conquest is fully operational file system that
  integrates persistent RAM with disk storage to
  provide significantly improved performance
• With RAM/disk hybrid, 43 to 96 speedup is
  gained compared to popular disk-based file system
• Benefits of Conquest arose from rethinking basic
  file system design assumptions