Hello, I am a modest user (no expert/developer) of a modest (old) desktop, and when it comes to how filesystems perform on _my_ desktop I've always had a very clear impression of those that work good (fast) and those that work bad (slow). In short, Ext3 and ReiserFS are the ones that always performed good, while all the rest, including JFS, XFS, Reiser4, FAT32, NTFS (on Windows), ZFS (on Solaris) and UFS (on FreeBSD), perform badly. What I've never been sure of is if this is because: a) My hardware is quite slow and therefor the difference in how each filesystem performs is much more noticeable b) This behavior is specific to my hardware, that for whatever reason it only likes Ext3 and Reiserfs (I used ReiserFS some years ago, but I use Ext3 now). I've always thought that the slowest operation on a HDD should normally be small random writes, and so this would be the main responsible for desktop users feeling that HDD are slow. Reading lately about netbooks (Eee PC and the like) that come with low end SSD and how many users complain about slowness and even freezes has somehow confirmed this theory for me, since these low end SSD are known to be quite slow at small random writes. So I've searched for a tool to measure this and I've come across Sysbench (and on Windows I've found one called Crystal DiskMark that can also measure this). The results have not surprised me much (please keep in mind that this is a 5 y.o Pentuim 4 with a 40Gb HD @5400, so figures are low): XFS: 81 Kb/s JFS: 93 Kb/s FAT32 (linux): 99 Kb/s NTFS (Windows): 108 Kb/s FAT32 (Windows): 110 Kb/s ReiserFS: 214 Kb/s Ext3: 218 Kb/s All filesystems were created with default options and mounted with noatime as the only option. They were all tested in the same last 5.2Gb partition of the disk, using a command like*: sysbench --test=fileio --file-block-size=4K --file-test-mode=rndwr prepare sysbench --test=fileio --file-block-size=4K --file-test-mode=rndwr run sysbench --test=fileio --file-block-size=4K --file-test-mode=rndwr cleanup *In fact I also used --file-total-size=100M so that the tests would complete faster than with the default 2G file size, but that made no difference in results. Also note that the test on Windows is not exactly the same, just a similar one. So Ext3 and ReiserFS are ~100%+ faster than all the rest in this slow operation, and that really makes a difference for me (I didn't test other filesystems mentioned above as slow, but I'm pretty sure they will perform similar to the rest of the "bad" group). Curious thing: I also remembered having tried Ext3 with data=writeback in the past and not liking it, even if it was supposed to be faster. So I tested Ext3 with the other two journal options: Ext3 writeback: 98 Kb/s Ext3 journal: 102 Kb/s So it is data=ordered that's making the huge difference in Ext3. Strange! Then out of curiosity I also tested Btrfs 0.16: Btrfs: 216 Kb/s Btrfs nodatacow: 235 Kb/s Looking good! I tried to test Ext4, but got an error on mounting about the partition not having a journal (or something similar) and didn't investigate any further. However, in my previous experiences with Ext4 it performed as good as Ext3 (or even slightly better). So, to finish up, this could be something specific to my hardware and therefor this mail just a waste of time (sorry, in that case), but then again, my hardware is quite standard -if old- (Intel chipset with integrated graphics, Seagate HDD, MSI motherboard,...) so maybe it's something common and worth looking at by people who "know"? Ah, tests were done using kernel 2.6.26.5, but this behavior has been the same for ages, so I don't think that matters. Thanks, Alberto. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
