Posted by hansivers on Fri 21 Apr 2006 at 11:10
There are a lot of Linux filesystems comparisons available but most of them are anecdotal, based on artificial tasks or completed under older kernels. This benchmark essay is based on 11 real-world tasks appropriate for a file server with older generation hardware (Pentium II/III, EIDE hard-drive).
Since its initial publication, this article has generated a lot of questions, comments and suggestions to improve it. Consequently, I'm currently working hard on a new batch of tests to answer as many questions as possible (within the original scope of the article). Results will be available in about two weeks (May 8, 2006) Many thanks for your interest and keep in touch with Debian-Administration.org! Hans
Why another benchmark test?
I found two quantitative and reproductible benchmark testing studies using the 2.6.x kernel (see References). Benoit (2003) implemented 12 tests using large files (1+ GB) on a Pentium II 500 server with 512MB RAM. This test was quite informative but results are beginning to aged (kernel 2.6.0) and mostly applied to settings which manipulate exclusively large files (e.g., multimedia, scientific, databases).
Piszcz (2006) implemented 21 tasks simulating a variety of file operations on a PIII-500 with 768MB RAM and a 400GB EIDE-133 hard disk. To date, this testing appears to be the most comprehensive work on the 2.6 kernel. However, since many tasks were "artificial" (e.g., copying and removing 10 000 empty directories, touching 10 000 files, splitting files recursively), it may be difficult to transfer some conclusions to real-world settings.
Thus, the objective of the present benchmark testing is to complete some Piszcz (2006) conclusions, by focusing exclusively on real-world operations found in small-business file servers (see Tasks description).
Description of selected tasks
The sequence of 11 tasks (from creation of FS to umounting FS) was run as a Bash script which was completed three times (the average is reported). Each sequence takes about 7 min. Time to complete task (in secs), percentage of CPU dedicated to task and number of major/minor page faults during task were computed by the GNU time utility (version 1.7).
Initial (after filesystem creation) and residual (after removal of all files) partition capacity was computed as the ratio of number of available blocks by number of blocks on the partition. Ext3 has the worst inital capacity (92.77%), while others FS preserve almost full partition capacity (ReiserFS = 99.83%, JFS = 99.82%, XFS = 99.95%). Interestingly, the residual capacity of Ext3 and ReiserFS was identical to the initial, while JFS and XFS lost about 0.02% of their partition capacity, suggesting that these FS can dynamically grow but do not completely return to their inital state (and size) after file removal.
Conclusion : To use the maximum of your partition capacity, choose ReiserFS, JFS or XFS.
File system creation, mounting and unmounting
The creation of FS on the 20GB test partition took 14.7 secs for Ext3, compared to 2 secs or less for other FS (ReiserFS = 2.2, JFS = 1.3, XFS = 0.7). However, the ReiserFS took 5 to 15 times longer to mount the FS (2.3 secs) when compared to other FS (Ext3 = 0.2, JFS = 0.2, XFS = 0.5), and also 2 times longer to umount the FS (0.4 sec). All FS took comparable amounts of CPU to create FS (between 59% - ReiserFS and 74% - JFS) and to mount FS (between 6 and 9%). However, Ex3 and XFS took about 2 times more CPU to umount (37% and 45%), compared to ReiserFS and JFS (14% and 27%).
Conclusion : For quick FS creation and mounting/unmounting, choose JFS or XFS.
Operations on a large file (ISO image, 700MB)
The initial copy of the large file took longer on Ext3 (38.2 secs) and ReiserFS (41.8) when compared to JFS and XFS (35.1 and 34.8). The recopy on the same disk advantaged the XFS (33.1 secs), when compared to other FS (Ext3 = 37.3, JFS = 39.4, ReiserFS = 43.9). The ISO removal was about 100 times faster on JFS and XFS (0.02 sec for both), compared to 1.5 sec for ReiserFS and 2.5 sec for Ext3! All FS took comparable amounts of CPU to copy (between 46 and 51%) and to recopy ISO (between 38% to 50%). The ReiserFS used 49% of CPU to remove ISO, when other FS used about 10%. There was a clear trend of JFS to use less CPU than any other FS (about 5 to 10% less). The number of minor page faults was quite similar between FS (ranging from 600 - XFS to 661 - ReiserFS).
Conclusion : For quick operations on large files, choose JFS or XFS. If you need to minimize CPU usage, prefer JFS.
Operations on a file tree (7500 files, 900 directories, 1.9GB)
The initial copy of the tree was quicker for Ext3 (158.3 secs) and XFS (166.1) when compared to ReiserFS and JFS (172.1 and 180.1). Similar results were observed during the recopy on the same disk, which advantaged the Ext3 (120 secs) compared to other FS (XFS = 135.2, ReiserFS = 136.9 and JFS = 151). However, the tree removal was about 2 times longer for Ext3 (22 secs) when compared to ReiserFS (8.2 secs), XFS (10.5 secs) and JFS (12.5 secs)! All FS took comparable amounts of CPU to copy (between 27 and 36%) and to recopy the file tree (between 29% - JFS and 45% - ReiserFS). Surprisingly, the ReiserFS and the XFS used significantly more CPU to remove file tree (86% and 65%) when other FS used about 15% (Ext3 and JFS). Again, there was a clear trend of JFS to use less CPU than any other FS. The number of minor page faults was significantly higher for ReiserFS (total = 5843) when compared to other FS (1400 to 1490). This difference appears to come from a higher rate (5 to 20 times) of page faults for ReiserFS in recopy and removal of file tree.
Conclusion : For quick operations on large file tree, choose Ext3 or XFS. Benchmarks from other authors have supported the use of ReiserFS for operations on large number of small files. However, the present results on a tree comprising thousands of files of various size (10KB to 5MB) suggest than Ext3 or XFS may be more appropriate for real-world file server operations. Even if JFS minimize CPU usage, it should be noted that this FS comes with significantly higher latency for large file tree operations.
Directory listing and file search into the previous file tree
The complete (recursive) directory listing of the tree was quicker for ReiserFS (1.4 secs) and XFS (1.8) when compared to Ext3 and JFS (2.5 and 3.1). Similar results were observed during the file search, where ReiserFS (0.8 sec) and XFS (2.8) yielded quicker results compared to Ext3 (4.6 secs) and JFS (5 secs). Ext3 and JFS took comparable amounts of CPU for directory listing (35%) and file search (6%). XFS took more CPU for directory listing (70%) but comparable amount for file search (10%). ReiserFS appears to be the most CPU-intensive FS, with 71% for directory listing and 36% for file search. Again, the number of minor page faults was 3 times higher for ReiserFS (total = 1991) when compared to other FS (704 to 712).
Conclusion : Results suggest that, for these tasks, filesystems can be regrouped as (a) quick and more CPU-intensive (ReiserFS and XFS) or (b) slower but less CPU-intensive (ext3 and JFS). XFS appears as a good compromise, with relatively quick results, moderate usage of CPU and acceptable rate of page faults.
These results replicate previous observations from Piszcz (2006) about reduced disk capacity of Ext3, longer mount time of ReiserFS and longer FS creation of Ext3. Moreover, like this report, both reviews have observed that JFS is the lowest CPU-usage FS. Finally, this report appeared to be the first to show the high page faults activity of ReiserFS on most usual file operations.
While recognizing the relative merits of each filesystem, only one filesystem can be install for each partition/disk. Based on all testing done for this benchmark essay, XFS appears to be the most appropriate filesystem to install on a file server for home or small-business needs :
While Piszcz (2006) did not explicitly recommand XFS, he concludes that "Personally, I still choose XFS for filesystem performance and scalability". I can only support this conclusion.
Benoit, M. (2003). Linux File System Benchmarks.
Piszcz, J. (2006). Benchmarking Filesystems Part II. Linux Gazette, 122 (January 2006).