On Wed, 23 Sep 2020, Dave Chinner wrote: > > > dir-test /mnt/test/linux-2.6 63000 1048576 > > > nvfs 6.6s > > > ext4 dax 8.4s > > > xfs dax 12.2s > > > > > > > > > dir-test /mnt/test/linux-2.6 63000 1048576 link > > > nvfs 4.7s > > > ext4 dax 5.6s > > > xfs dax 7.8s > > > > > > dir-test /mnt/test/linux-2.6 63000 1048576 dir > > > nvfs 8.2s > > > ext4 dax 15.1s > > > xfs dax 11.8s > > > > > > Yes, nvfs is faster than both ext4 and XFS on DAX, but it's not a > > > huge difference - it's not orders of magnitude faster. > > > > If I increase the size of the test directory, NVFS is order of magnitude > > faster: > > > > time dir-test /mnt/test/ 2000000 2000000 > > NVFS: 0m29,395s > > XFS: 1m59,523s > > EXT4: 1m14,176s > > What happened to NVFS there? The runtime went up by a factor of 5, > even though the number of ops performed only doubled. This test is from a different machine (K10 Opteron) than the above test (Skylake Xeon). I borrowed the Xeon for a short time and I no longer have access to it. > > time dir-test /mnt/test/ 8000000 8000000 > > NVFS: 2m13,507s > > XFS: 14m31,261s > > EXT4: reports "file 1976882 can't be created: No space left on device", > > (although there are free blocks and inodes) > > Is it a bug or expected behavior? > > Exponential increase in runtime for a workload like this indicates > the XFS journal is too small to run large scale operations. I'm > guessing you're just testing on a small device? In this test, the pmem device had 64GiB. I've created 1TiB ramdisk, formatted it with XFS and ran dir-test 8000000 on it, however it wasn't much better - it took 14m8,824s. > In which case, you'd get a 16MB log for XFS, which is tiny and most > definitely will limit performance of any large scale metadta > operation. Performance should improve significantly for large scale > operations with a much larger log, and that should bring the XFS > runtimes down significantly. Is there some mkfs.xfs option that can increase log size? > > If you think that the lack of journaling is show-stopper, I can implement > > it. > > I did not say that. My comments are about the requirement for > atomicity of object changes, not journalling. Journalling is an > -implementation that can provide change atomicity-, it is not a > design constraint for metadata modification algorithms. > > Really, you can chose how to do object update however you want. What > I want to review is the design documentation and a correctness proof > for whatever mechanism you choose to use. Without that information, > we have absolutely no chance of reviewing the filesystem > implementation for correctness. We don't need a proof for something > that uses journalling (because we all know how that works), but for > something that uses soft updates we most definitely need the proof > of correctness for the update algorithm before we can determine if > the implementation is good... > > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx I am thinking about this: I can implement lightweight journaling that will journal just a few writes - I'll allocate some small per-cpu intent log for that. For example, in nvfs_rename, we call nvfs_delete_de and nvfs_finish_add - these functions are very simple, both of them write just one word - so we can add these two words to the intent log. The same for setattr requesting simultaneous uid/gid/mode change - they are small, so they'll fit into the intent log well. Regarding verifiability, I can do this - the writes to pmem are wrapped in a macro nv_store. So, I can modify this macro so that it logs all modifications. Then I take the log, cut it at random time, reorder the entries (to simulate reordering in the CPU write-combining buffers), replay it, run nvfsck on it and mount it. This way, we can verify that no matter where the crash happened, either an old file or a new file is present in a directory. Do you agree with that? Mikulas
