On 28.02.2020 18:35, Andreas Dilger wrote: > On Feb 27, 2020, at 5:24 AM, Kirill Tkhai <ktkhai@xxxxxxxxxxxxx> wrote: >> >> On 27.02.2020 00:51, Andreas Dilger wrote: >>> On Feb 26, 2020, at 1:05 PM, Kirill Tkhai <ktkhai@xxxxxxxxxxxxx> wrote: >>>> Why? There are two contradictory actions that filesystem can't do at the same time: >>>> >>>> 1)place files on a distance from each other to minimize number of extents >>>> on possible future growth; >>>> 2)place small files in the same big block of block device. >>>> >>>> At initial allocation time you never know, which file will stop grow in some >>>> future, i.e. which file is suitable for compaction. This knowledge becomes >>>> available some time later. Say, if a file has not been changed for a month, >>>> it is suitable for compaction with another files like it. >>>> >>>> If at allocation time you can determine a file, which won't grow in the future, >>>> don't be afraid, and just share your algorithm here. >>> >>> Very few files grow after they are initially written/closed. Those that >>> do are almost always opened with O_APPEND (e.g. log files). It would be >>> reasonable to have O_APPEND cause the filesystem to reserve blocks (in >>> memory at least, maybe some small amount on disk like 1/4 of the current >>> file size) for the file to grow after it is closed. We might use the >>> same heuristic for directories that grow long after initial creation. >> >> 1)Lets see on a real example. I created a new ext4 and started the test below: >> https://gist.github.com/tkhai/afd8458c0a3cc082a1230370c7d89c99 >> >> Here are two files written. One file is 4Kb. One file is 1Mb-4Kb. >> >> $filefrag -e test1.tmp test2.tmp >> Filesystem type is: ef53 >> File size of test1.tmp is 4096 (1 block of 4096 bytes) >> ext: logical_offset: physical_offset: length: expected: flags: >> 0: 0.. 0: 33793.. 33793: 1: last,eof >> test1.tmp: 1 extent found >> File size of test2.tmp is 1044480 (255 blocks of 4096 bytes) >> ext: logical_offset: physical_offset: length: expected: flags: >> 0: 0.. 254: 33536.. 33790: 255: last,eof >> test2.tmp: 1 extent found > > The alignment of blocks in the filesystem is much easier to see if you use > "filefrag -e -x ..." to print the values in hex. In this case, 33536 = 0x8300 > so it is properly aligned on disk IMHO. > >> $debugfs: testb 33791 >> Block 33791 not in use >> >> test2.tmp started from 131Mb. In case of discard granuality is 1Mb, test1.tmp >> placement prevents us from discarding next 1Mb block. > > For most filesystem uses, aligning the almost 1MB file on a 1MB boundary > is good. That allows a full-stripe read/write for RAID, and is more > likely to align with the erase block for flash. If it were to be allocated > after the 4KB block, then it may be that each 1MB-aligned read/write of a > large file would need to read/write two unaligned chunks per syscall. > >> 2)Another example. Let write two files: 1Mb-4Kb and 1Mb+4Kb: >> >> # filefrag -e test3.tmp test4.tmp >> Filesystem type is: ef53 >> File size of test3.tmp is 1052672 (257 blocks of 4096 bytes) >> ext: logical_offset: physical_offset: length: expected: flags: >> 0: 0.. 256: 35840.. 36096: 257: last,eof >> test3.tmp: 1 extent found >> File size of test4.tmp is 1044480 (255 blocks of 4096 bytes) >> ext: logical_offset: physical_offset: length: expected: flags: >> 0: 0.. 254: 35072.. 35326: 255: last,eof >> test4.tmp: 1 extent found > > Here again, "filefrag -e -x" would be helpful. 35840 = 0x8c00, and > 35072 = 0x8900, so IMHO they are allocated properly for most uses. > Packing all files together sequentially on disk is what FAT did and > it always got very fragmented in the end. > >> They don't go sequentially, and here is fragmentation starts. >> >> After both the tests: >> $df -h >> /dev/loop0 2.0G 11M 1.8G 1% /root/mnt >> >> Filesystem is free, all last block groups are free. E.g., >> >> Group 15: (Blocks 491520-524287) csum 0x3ef5 [INODE_UNINIT, ITABLE_ZEROED] >> Block bitmap at 272 (bg #0 + 272), csum 0xd52c1f66 >> Inode bitmap at 288 (bg #0 + 288), csum 0x00000000 >> Inode table at 7969-8480 (bg #0 + 7969) >> 32768 free blocks, 8192 free inodes, 0 directories, 8192 unused inodes >> Free blocks: 491520-524287 >> Free inodes: 122881-131072 >> >> but two files are not packed together. >> >> So, ext4 block allocator does not work good for my workload. It even does not >> know anything about discard granuality of underlining block device. Does it? >> I assume no fs knows. Should I tell it? > > You can tune the alignment of allocations via s_raid_stripe and s_raid_stride > in the ext4 superblock. I believe these are also set by mke2fs by libdisk, > but I don't know if it takes flash erase block geometry into account. > >>> The main exception there is VM images, because they are not really "files" >>> in the normal sense, but containers aggregating a lot of different files, >>> each created with patterns that are not visible to the VM host. In that >>> case, it would be better to have the VM host tell the filesystem that the >>> IO pattern is "random" and not try to optimize until the VM is cold. >>> >>>> In Virtuozzo we tried to compact ext4 with existing kernel interface: >>>> >>>> https://github.com/dmonakhov/e2fsprogs/blob/e4defrag2/misc/e4defrag2.c >>>> >>>> But it does not work well in many situations, and the main problem is blocks allocation in desired place is not possible. Block allocator can't behave >>>> excellent for everything. >>>> >>>> If this interface bad, can you suggest another interface to make block >>>> allocator to know the behavior expected from him in this specific case? >>> >>> In ext4 there is already the "group" allocator, which combines multiple >>> small files together into a single preallocation group, so that the IO >>> to disk is large/contiguous. The theory is that files written at the >>> same time will have similar lifespans, but that isn't always true. >>> >>> If the files are large and still being written, the allocator will reserve >>> additional blocks (default 8MB I think) on the expectation that it will >>> continue to write until it is closed. >>> >>> I think (correct me if I'm wrong) that your issue is with defragmenting >>> small files to free up contiguous space in the filesystem? I think once >>> the free space is freed of small files that defragmenting large files is >>> easily done. Anything with more than 8-16MB extents will max out most >>> storage anyway (seek rate * IO size). >> >> My issue is mostly with files < 1Mb, because underlining device discard >> granuality is 1Mb. The result of fragmentation is that size of occupied >> 1Mb blocks of device is 1.5 times bigger, than size of really written >> data (say, df -h). And this is the problem. > > > Sure, and the group allocator will aggregate writes << prealloc size of > 8MB by default. If it is 1MB that doesn't qualify for group prealloc. > I think under 64KB does qualify for aggregation and unaligned writes. > >>> In that case, an interesting userspace interface would be an array of >>> inode numbers (64-bit please) that should be packed together densely in >>> the order they are provided (maybe a flag for that). That allows the >>> filesystem the freedom to find the physical blocks for the allocation, >>> while userspace can tell which files are related to each other. >> >> So, this interface is 3-in-1: >> >> 1)finds a placement for inodes extents; > > The target allocation size would be sum(size of inodes), which should > be relatively small in your case). > >> 2)assigns this space to some temporary donor inode; > > Maybe yes, or just reserves that space from being allocated by anyone. > >> 3)calls ext4_move_extents() for each of them. > > ... using the target space that was reserved earlier > >> Do I understand you right? > > Correct. That is my "5 minutes thinking about an interface for grouping > small files together without exposing kernel internals" proposal for this. Ok. I'll think about the prototype and then public to the mailing list. >> If so, then IMO it's good to start from two inodes, because here may code >> a very difficult algorithm of placement of many inodes, which may require >> much memory. Is this OK? > > Well, if the files are small then it won't be a lot of memory. Even so, > the kernel would only need to copy a few MB at a time in order to get > any decent performance, so I don't think that is a huge problem to have > several MB of dirty data in flight. I mean not in-flight IO, but memory for all logic of files placement. Userspace may build multi-step algoritm, which is hidden for kernel: pack two files together, then decrease number of extents of some third file, then pack something else. Also, files related to different directories should be packed together, but it does not look good for kernel to look for files directories by inodes (our interface is about 64-bit inodes numbers, sure?). For me it does not look good, kernel iterates over all files and looks for a placement for a specific file, since this is just excess work for kernel. Usually, both the files are chosen by userspace, and the userspace does not want to move more then one of them at time. >> Can we introduce a flag, that some of inode is unmovable? > > There are very few flags left in the ext4_inode->i_flags for use. > You could use "IMMUTABLE" or "APPEND_ONLY" to mean that, but they > also have other semantics. The EXT4_NOTAIL_FL is for not merging the > tail of a file, but ext4 doesn't have tails (that was in Reiserfs), > so we might consider it a generic "do not merge" flag if set? > >> Can this interface use a knowledge about underlining device discard granuality? > > As I wrote above, ext4+mballoc has a very good appreciation for alignment. > That was written for RAID storage devices, but it doesn't matter what > the reason is. It isn't clear if flash discard alignment is easily > used (it may not be a power-of-two value or similar), but wouldn't be > harmful to try. > >> In the answer to Dave, I wrote a proposition to make fallocate() care about >> i_write_hint. Could you please comment what you think about that too? > > I'm not against that. How the two interact would need to be documented > first and discussed to see if that makes sene, and then implemented. Thanks, Kirill
