Hi Mikulas, I'll say up front that I think you're barking up the wrong tree trying to knock down XFS and ext4 to justify NVFS. NVFS will stand or fall on it's own merits, not on how you think it's better than other filesystems... I have some fundamental concerns about the NVFS integrity model, they came out as I wrote this response to your characterisations of XFS and journalling filesysetms. Maybe I'm missing something about NVFS that isn't clearly explained.... On Mon, Sep 21, 2020 at 12:20:42PM -0400, Mikulas Patocka wrote: > On Wed, 16 Sep 2020, Mikulas Patocka wrote: > > On Wed, 16 Sep 2020, Dan Williams wrote: > > > On Wed, Sep 16, 2020 at 10:24 AM Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: > > > > > > > > > My first question about nvfs is how it compares to a daxfs with > > > > > executables and other binaries configured to use page cache with the > > > > > new per-file dax facility? > > > > > > > > nvfs is faster than dax-based filesystems on metadata-heavy operations > > > > because it doesn't have the overhead of the buffer cache and bios. See > > > > this: http://people.redhat.com/~mpatocka/nvfs/BENCHMARKS > > > > > > ...and that metadata problem is intractable upstream? Christoph poked > > > at bypassing the block layer for xfs metadata operations [1], I just > > > have not had time to carry that further. > > > > > > [1]: "xfs: use dax_direct_access for log writes", although it seems > > > he's dropped that branch from his xfs.git > > > > XFS is very big. I wanted to create something small. > > And the another difference is that XFS metadata are optimized for disks > and SSDs. Ah, that old chestnut. :) > On disks and SSDs, reading one byte is as costly as reading a full block. > So we must put as much information to a block as possible. XFS uses > b+trees for file block mapping and for directories - it is reasonable > decision because b+trees minimize the number of disk accesses. Actually, no, that wasn't why XFS was implemented using btrees. The btrees are an implementation detail, not a design requirement to minimise the number of disk accesses. XFS was intended for huge disk arrays (hundreds to thousands on individual spindles) and so no attempt was made in the design to minimise disk accesses. There was -always- going to be a huge number of IOPS available in the intended large scale deployments, so concurrency and *efficiency at scale* was far more important at the design level than minimising the number of disk ops for any given operation. To that end, simulations were done that showed that extent based trees were much more CPU, memory and IO efficient than bitmaps, hybrid tree-bitmaps or multi-layer indirect block referencing when trying to index and access large amounts of data. To be efficient at scale, all operations need to be O(log N) or better, and extent based encoding is much, more compact than direct block indexing. Extent trees are also much more effective for finding exact fits, identifying large contiguous spaces, and manipulating large ranges of indexed data than other filesystem indexing mechanisms. They are also not bound by alignment restrictions like hierarchical binary/buddy bitmap schemes are, and their maximum size is bounded and can be calculated at runtime. IOWs, extent based trees were chosen because of scalability, efficiency, and flexibility reasons before the actual tree structure that it would be implemented with was decided on. b+trees were used in the implementation because one tree implementation could do everything as all that needed to change btree trees was the pointer and record format. The result of this is that we have made -zero- changes to the XFS structure and algorithms for SSDs. We don't do different things based on the blkdev rotational flag, or anything like that. XFS behaves exactly the same on spinning disks as it does SSDs as it does PMEM and it performs well on all of them. And that performance doesn't drop away as you increase the scale and capability of the underlying storage. That's what happens when storage algorithms are designed for concurrency and efficiency at scale rather than optimising for a specific storage characteristic. NVFS is optimised for a specific storage characteristic (i.e. low latency synchronous storage), so I would absolutely expect it to be faster than XFS on that specific storage. However, claims like this: > On persistent memory, each access has its own cost, so NVFS uses metadata > structures that minimize the number of cache lines accessed (rather than > the number of blocks accessed). For block mapping, NVFS uses the classic > unix dierct/indirect blocks - if a file block is mapped by a 3-rd level > indirect block, we do just three memory accesses and we are done. If we > used b+trees, the number of accesses would be much larger than 3 (we would > have to do binary search in the b+tree nodes). ... are kinda naive, because you're clearly optimising the wrong aspect of block mapping. Extents solve the block indexing overhead problem; optimising the type of tree you use to index the indirect blocks doesn't avoid the overhead of having to iterate every block for range operations. IOWs, we use extents because they are space and time efficient for the general use cases. XFS can map 2^21 blocks into a single 16 byte extent record (8GiB file mapping for 4k block size) and so the vast majority of files in a filesystem are mapped with a single extent. The NVFS indirect block tree has a fan-out of 16, mapping 2^21 blocks requires a 5 level indirect tree. Whcih one if going to be faster to truncate away - a single record or 2 million individual blocks? IOWs, we can take afford to take an extra cacheline miss or two on a tree block search, because we're accessing and managing orders of magnitude fewer records in the mapping tree than an indirect block tree. PMEM doesn't change this: extents are more time and space efficient at scale for mapping trees than indirect block trees regardless of the storage medium in use. > The same for directories - NVFS hashes the file name and uses radix-tree > to locate a directory page where the directory entry is located. XFS > b+trees would result in much more accesses than the radix-tree. That's like me saying that XFS hashes the file name and uses a btree to index the directory block where the dirent is located, so it will be faster than a radix tree. It's completely bogus. It ignores the fact that both filesysetms use the same hash based lookup indexing algorithms and use O(log N) trees for the name hash. IOWs, the only real difference is the fan-out and depths of the tree. The end result is that algorithmic performance of name -> dirent lookups are going to be *very similar* and, as such, the performance differential is going to be dominated by other implementation differences. Such as the fact that XFS has to buffer the directory metadata, hence that the initial directory block lookup cost is higher than NVFS. Subsequent block lookups hit the buffer cache, so that caching overhead is somewhat amortised over multiple directory accesses, but it doesn't get rid of it. IOWs, difference in memory accesses between a radix tree and btree for this algorithm is largely irrelevant, and even your tests indicate that. The modification tests show that metadata lookup *and journalling* overhead is not really that significant as the number of directory entries increase: 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 XFS and ext4 directory structures were substantially less efficient than the NVFS structure, then NVFS should absolutely -slaughter- ext4 and XFS in directory modification intensive microbenchmarks like this. Especially considering that difference in performance also includes the journalling overhead. IOWs, the differences in performance are not a result of the directory structures or the number of memory fetches their indexing algorithms require to do the work - the differences come from structural features: ext4 and XFS have more work to do per directory operation thanks to their metadata buffer and journalling management requirements. Also, keep in mind that much of the complexity in the XFS directory structure doesn't make XFS go faster at small directory sizes. They actually slow it down at small sizes, but they also stop performance from falling off a cliff at scale. Hence results might be quite different if you are running with millions of directory entries in the directories rather that a few thousand.... > Regarding journaling - NVFS doesn't do it because persistent memory is so > fast that we can just check it in the case of crash. NVFS has a > multithreaded fsck that can do 3 million inodes per second. Scanning speed means little when it comes to integrity checking. Fast storage can hide a multitude of sins, the least of which is inefficient algorithms. More importantly, it can hide crash related inconsistencies, because timing the crash to land between two specific modifications is much harder on fast storage than it is slow storage. IOWs, just because fsck can iterate inodes at 3M a second doesn't mean the filesystem code is crash safe and correct, nor that fsck can detect and correct all the inconsistencies that the crash left behind and need fixing. More on this later.... > XFS does > journaling (it was reasonable decision for disks where fsck took hours) > and it will cause overhead for all the filesystem operations. Fundamentally, journalling provides guarantees much more important than than "does not need fsck". Journalling provides -atomic metadata changes-, and that's something you can't do without some variant of journalled, log structured or COW metadata. This is important, because atomicity of metadata changes is something users actually expect from filesystems. Take, for example, truncate. If you punch out the space on storage before you change the inode size on storage and then crash in-between the punch and the inode size reduction, the user file is left with a bunch of zeros in it over the range between the new EOF and the old EOF. Users will see partial completion state. IOWs, the NVFS truncate operation as implemented: if (attr->ia_valid & ATTR_SIZE) { WARN_ON(!inode_is_locked(&nmi->vfs_inode)); if (attr->ia_size != nmi->vfs_inode.i_size) { r = nvfs_punch_hole(nmi, attr->ia_size, LLONG_MAX - ((1UL << nvs->log2_page_size) - 1)); if (unlikely(r)) return r; nvfs_set_inode_size(nmi, attr->ia_size); } } is not atomic from a user crash recovery perspective as it exposes partially complete state to the user. For it to be atomic from the user perspective, on truncate down the inode size must be changed on disk first, then the space beyond the new EOF needs to get punched out. hence if we crash while the punching is occurring, users will not see that after remount because the inconsistent state is beyond the range they can access. IOWs, they see the file as if the truncate down is fully complete, regardless of whether it is actually complete or not. However, that then potentially breaks truncate up because, conversely, truncate up requires that any blocks already allocated beyond EOF needs to be zeroed before the inode size is changed so stale data is not exposed between the old EOF and the new EOF.... Yes, this can be fixed by changing the order of the punch vs the inode size change based on what type of operation is actually going to be performed, but this is just an example of the complexity problem "soft updates" bring to otherwise "simple" operations. I haven't even mentioned freeing indirect blocks also updates free space, inode i_blocks counters, potentially multiple indirect blocks, etc. If the order is wrong, then all bets are off for what fsck will actually do with the inode when it scans it and finds partially complete state. And even if it gets corrected, there's no guarantee taht the entire operation was completed from the perspective of the user. Rename is another operation that has specific "operation has atomic behaviour" expectations. I haven't looked at how you've implementated that yet, but I suspect it also is extremely difficult to implement in an atomic manner using direct pmem updates to the directory structures. AFAICS, it is easy to get metadata update ordering wrong, and without a formal proof that every single write to storage is correctly ordered I can't see how this model can be validated and reviewed. It is made exceedingly complex by direct storage access. instead of doing all the changes to a single block in memory and then only having to order block writes to stable storage correctly (as per the BSD filesystem that used soft updates), every direct pmem modification to an object needs to be ordered correctly against all other direct pmem modifications, both within the object and across objects. And this brings me back to modification atomicity - soft update ordering across objects is hard enough, but how do you perform dependent modifications to a single object atomically? e.g. Looking at nvfs_setattr, why is it correct to write timestamp changes to pmem before the truncate is done? And if that is correct behavour, then why is it correct to change UID/GID _before_ updating the timestamp? And if that is also correct behaviour, then why are mode changes done _after both_ uid/gid and timestamp changes? What happens if setattr is asked to change ATTR_UID|ATTR_GID|ATTR_MODE as an atomic change and we crash before the mode has been changed? I think the only right answer can be "->setattr changes are atomic at the stable storage level", in which case the way NVFS is updating stable storage breaking all sorts of assumptions and expectations for changing security and ownership attributes of the inode. And to bring this right back to "fsck is fast so we don't need journalling": how would running fsck before mounting detect that these compound object updates were not fully completed? How does running fsck after the crash guarantee compound object modifications it knows nothing about are executed atomically? That, too me, looks like a fundamental, unfixable flaw in this approach... I can see how "almost in place" modification can be done by having two copies side by side and updating one while the other is the active copy and switching atomically between the two objects. That way a traditional soft-update algorithm would work because the exposure of the changes is via ordering the active copy switches. That would come at a cost, though, both in metadata footprint and CPU overhead. So, what have I missed about the way metadata is updated in the pmem that allows non-atomic updates to work reliably? Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
