JFS Tuning and Performance

19
Table 3. Default intent log size
FS Size
VxFS 3.5 or 4.1 or DLV <= 5
VxFS 5.0 or later and DLV >= 6
<= 8 MB
1 MB
1 MB
<= 512 MB
16 MB
16 MB
<= 16 GB
16 MB
64 MB
> 16 GB
16 MB
256 MB
For most applications, the default log size is sufficient. However, large file systems with heavy
structural changes simultaneously by multiple threads or heavy synchronous write operations with
datainlog may need a larger Intent Log to prevent transaction stalls when the log is full.
Disk layout version
As mentioned earlier, several improvements have been introduced in the disk layouts which can help
improve performance. The supported and default disk layout version will depend on the operating
system version and the VxFS version. Remember that later disk layout versions often support new
performance enhancements, such as faster performance for large directories available with disk
layout version 5.
See the manpages mkfs_vxfs(1M) and newfs_vxfs(1M) for more information on file system create
options.
Large directories
One myth about fragmentation is that a file system with a large number of small files (for example
500,000 files which are 4 KB in size or less) does not need to be defragmented. However, a file
system with thousands of small files is likely to have very large directories. Directories are good
examples of files that are often opened, extended, and closed. Therefore, directories are usually
fragmented.
“Small” is a relative term. The number of blocks taken by a directory also depends on the size of each
file name. As a general rule, consider a small directory to be one that has fewer than 10,000
directory entries.
When adding a new file to a directory or looking for a non-existent file, every directory block must be
searched. If the directory has 1000 directory blocks, then the system must do at least 1000 I/Os to
add a single file to the directory or search for a non-existent file.
Simultaneous directory searches also incur contention on the inode, extent map, or directory blocks,
potentially single-threading access to the directory. Long delays can be detected when doing multiple
„ll‟ commands on a single large directory.
Large directories can be defragmented when doing extent reorganizations, so that the directory
contains larger but fewer extents. The reorganization of a directory can relieve bottlenecks on the
indirect blocks when large directories are searched simultaneously.
VxFS 5.0 on 11i v2 and 11i v3 introduced a new indexing mechanism which avoids the sequential
scanning of directory blocks. To take advantage of this feature, the file system must be created or
upgraded to use disk layout version 7.