[summit]2009 Red Hat Enterprise Linux File Systems들에 대한 구조/성능 비교 (ext4/xfs)

[이누기]

Red Hat Enterprise Linux File Systems

파일시스템 ext2, ext3 , XFS, ext4, Btrfs 에 대한 비교와 구조 설명

성능 비교/파일시스템의 제한등의 내용을 상세히 담고 있습니다.

아래는 Ext4에 대한 특징들입니다. 

Ext4

● 개발자들은 Mingming Cao, Andreas Dilger, Alex Tomas, Dave Kleikamp, Theodore Ts'o, Eric Sandeen, Sam Naghshineh,  기타  입니다.

● Red Hat Enterprise Linux 5.3 와5.4 에서 Tech preview 로 지원됩니다.

● ext3의 확장적인 측면으로 시작되었습니다.

- ext3보다 성능 확장성, 신뢰성의 이슈가 좋아졌습니다.

-  Fork of ext3 in 2006 to not affect current ext3 users

-  2008년도 2.6.28 커널에 안정적으로 추가되었습니다.

● Fedora 11 과 다른 배포판에 기본적으로 탑재되었습니다.

- Fedora 11 – first distribution shipping a testing version

- Red Hat Enterprise Linux 5.1 – first EL w/tech preview

● ext2 와 ext3의 호환성

- Can mount existing ext2 and ext3 file systems as ext4

- Can mount ext4 as ext3

  -  ... assuming extents were never used

● Delayed allocation

- Don't allocate blocks until data is going to be written to disk

- Improves performance

- Reduces fragmentation – allocations based upon actual file size

● Multiblock allocator

- Multiple blocks for a file allocated in a single operation

- Reduces fragmentation by choosing contiguous blocks

● Delayed and multiblock allocation observed benefits

-- Improved throughput by 30%

-- Reduced CPU usage by 50%

● Extents

- Replaces block mapping as used in ext2 and ext3

- Range of contiguous blocks

- Start/length pairs

- 215 blocks ? 128 MB with 4 KB block size

- Improves large file performance and reduces fragmentation

- Up to 4 extents can be stored in the inode, additional indexed in an Htree

- Extents w/48-bit block numbers (vs. 32-bit in ext3)...

  - 16 TB (ext3) ? 1 EB (ext4 with 4 KB block size) max file size

● Persistent file pre-allocation

- Old way – write a file full of 0s

- New way – reserve but don't take time to write out 0s

- Guaranteed available (unlike a sparse file)

- Likely contiguous

- Ideal for media streaming and databases

● 32,000 ? 64,000+ subdirectories in a directory

● Improved timestamps

- Second ? nanosecond granularity

- Year 2038 problem deferred to 2514

● Online defrag of individual files or entire file system

- e4defrag

- Currently in git and Fedora Rawhide

● Journal and block group checksums

● Faster fsck

- Problem: full fsck of 2 TB ext3 on high end RAID ? 2 to 4 hours to days

- Unallocated block groups and inodes are marked and don't need to be fsck'ed

- Consequence: 2x to 10x+ speed up

- Enabled by default or via -O uninit_groups

degts_1010_rhel_file_systems.pdf