HP StorageWorks Scalable File Share - Manual (Page 341)

Meta-data operations from a single client node B–13
B.9 Meta-data operations from a single client node
Figure B-11 shows the numbers of meta-data operations that are possible to a Lustre file system from a single
client node.
Figure B-11 Meta-data operations from a single client node.
The results shown in Figure B-11are based on the following configuration:
• File system with one OST service (that is, a file stripe of one) on a system using SFS20 storage
• Mirrored MDS service
The four tests were performed with an aggregate population of three million files, distributed as follows in
the tests:
• 30 directories, each containing 100,000 files
• 6 directories, each containing 500,000 files
• 3 directories, each containing 1 million files
• 2 directories, each containing 1.5 million files
• A single directory containing 3 million files
The results demonstrate that there is a correlation between the directory population and the aggregate
number of operations possible—that is, the smaller the individual directory population, the higher the
aggregate number of operations possible.
Note the following points:
•
mknod()
operations are faster than
creat()
operations because
mknod()
operations do not
require activity at the Object Storage Servers (whereas
creat()
operations create objects on the
OST service);
mknod()
operations represent the high-water mark of I/O operations.
• The rate at which files created by
mknod()
operations and files created by
creat()
operations can
be removed by the
unlink()
operation is also shown. This represents the high-water mark for file
removal; rates obtainable with the
rm -rf
command can be lower than this. Refer to Chapter 6 of
0
1000
2000
3000
4000
5000
6000
7000
Operations/sec
100000 500000 1 million 1.5 million 3 million
Directory Populations
HP SFS Meta-data Operations (3M files)
mknods
unlinks (mknod)
creats
unlinks (creat)
stats