146
9
F I L E S Y S T E M P E R F O R M A N C E
File system
Write (MB/sec and % of peak)
Read (MB/sec and % of peak)
BFS
5.88 (99%)
5.91 (99%)
ext2
4.59 (78%)
4.97 (84%)
NTFS
3.77 (64%)
3.12 (52%)
Table 9-2 IOZone bandwidths for a 128 MB file written in 64K chunks.
File system
Write (MB/sec and % of peak)
Read (MB/sec and % of peak)
BFS
5.88 (99%)
5.91 (99%)
ext2
4.36 (74%)
5.75 (97%)
NTFS
3.81 (64%)
3.05 (51%)
Table 9-3 IOZone bandwidths for a 128 MB file written in 256K chunks.
File system
Write (MB/sec and % of peak)
Read (MB/sec and % of peak)
BFS
5.81 (98%)
5.84 (98%)
ext2
4.31 (73%)
5.51 (93%)
NTFS
3.88 (65%)
3.10 (52%)
Table 9-4 IOZone bandwidths for a 512 MB file written in 128K chunks.
the disk (about 6 MB/sec) is significantly less than the
memcpy()
bandwidth
of the machine (approximately 50 MB/sec). If the disk subsystem were faster,
Linux would not perform as well relative to the maximum speed of the disk.
The BeOS approach to direct I/O works exceptionally well in this situation
and scales to higher-performance disk subsystems.
File Creation/Deletion Benchmark
The lmbench test suite by Larry McVoy and Carl Staelin is an extensive
benchmark suite that encompasses many areas of performance. One of the
tests from that suite,
lat fs
, tests the speed of create and delete operations
on a file system. Although highly synthetic, this benchmark provides an easy
yardstick for the cost of file creation and deletion.
We used the systems described previously for these tests. We also ran the
benchmark on a BFS volume created with indexing turned off. Observing the
speed difference between indexed and nonindexed BFS gives an idea of the
cost of maintaining the default indices (name, size, and last modified time).
The nonindexed BFS case is also a fairer comparison with NTFS and XFS
since they do not index anything.
We used
lat fs
v1.6 from the original lmbench test suite (not lmbench 2.0)
because it was easier to port to NT. The
lat fs
test creates 1000 files (writing
Practical File System Design:The Be File System
, Dominic Giampaolo
page 146