His description of the problem and solution follow. My own tests show
speedups on typical filesystem intensive workloads of 5% to 12% which
is very impressive considering the small amount of code change involved.
------
One day I noticed that some file operations run much faster on
small file systems then on big ones. I've looked at the ffs
algorithms, thought about them, and redesigned the dirpref algorithm.
First I want to describe the results of my tests. These results are old
and I have improved the algorithm after these tests were done. Nevertheless
they show how big the perfomance speedup may be. I have done two file/directory
intensive tests on a two OpenBSD systems with old and new dirpref algorithm.
The first test is "tar -xzf ports.tar.gz", the second is "rm -rf ports".
The ports.tar.gz file is the ports collection from the OpenBSD 2.8 release.
It contains 6596 directories and 13868 files. The test systems are:
1. Celeron-450, 128Mb, two IDE drives, the system at wd0, file system for
test is at wd1. Size of test file system is 8 Gb, number of cg=991,
size of cg is 8m, block size = 8k, fragment size = 1k OpenBSD-current
from Dec 2000 with BUFCACHEPERCENT=35
2. PIII-600, 128Mb, two IBM DTLA-307045 IDE drives at i815e, the system
at wd0, file system for test is at wd1. Size of test file system is 40 Gb,
number of cg=5324, size of cg is 8m, block size = 8k, fragment size = 1k
OpenBSD-current from Dec 2000 with BUFCACHEPERCENT=50
You can get more info about the test systems and methods at:
http://www.ptci.ru/gluk/dirpref/old/dirpref.html
Test Results
tar -xzf ports.tar.gz rm -rf ports
mode old dirpref new dirpref speedup old dirprefnew dirpref speedup
First system
normal 667 472 1.41 477 331 1.44
async 285 144 1.98 130 14 9.29
sync 768 616 1.25 477 334 1.43
softdep 413 252 1.64 241 38 6.34
Second system
normal 329 81 4.06 263.5 93.5 2.81
async 302 25.7 11.75 112 2.26 49.56
sync 281 57.0 4.93 263 90.5 2.9
softdep 341 40.6 8.4 284 4.76 59.66
"old dirpref" and "new dirpref" columns give a test time in seconds.
speedup - speed increasement in times, ie. old dirpref / new dirpref.
------
Algorithm description
The old dirpref algorithm is described in comments:
/*
* Find a cylinder to place a directory.
*
* The policy implemented by this algorithm is to select from
* among those cylinder groups with above the average number of
* free inodes, the one with the smallest number of directories.
*/
A new directory is allocated in a different cylinder groups than its
parent directory resulting in a directory tree that is spreaded across
all the cylinder groups. This spreading out results in a non-optimal
access to the directories and files. When we have a small filesystem
it is not a problem but when the filesystem is big then perfomance
degradation becomes very apparent.
What I mean by a big file system ?
1. A big filesystem is a filesystem which occupy 20-30 or more percent
of total drive space, i.e. first and last cylinder are physically
located relatively far from each other.
2. It has a relatively large number of cylinder groups, for example
more cylinder groups than 50% of the buffers in the buffer cache.
The first results in long access times, while the second results in
many buffers being used by metadata operations. Such operations use
cylinder group blocks and on-disk inode blocks. The cylinder group
block (fs->fs_cblkno) contains struct cg, inode and block bit maps.
It is 2k in size for the default filesystem parameters. If new and
parent directories are located in different cylinder groups then the
system performs more input/output operations and uses more buffers.
On filesystems with many cylinder groups, lots of cache buffers are
used for metadata operations.
My solution for this problem is very simple. I allocate many directories
in one cylinder group. I also do some things, so that the new allocation
method does not cause excessive fragmentation and all directory inodes
will not be located at a location far from its file's inodes and data.
The algorithm is:
/*
* Find a cylinder group to place a directory.
*
* The policy implemented by this algorithm is to allocate a
* directory inode in the same cylinder group as its parent
* directory, but also to reserve space for its files inodes
* and data. Restrict the number of directories which may be
* allocated one after another in the same cylinder group
* without intervening allocation of files.
*
* If we allocate a first level directory then force allocation
* in another cylinder group.
*/
My early versions of dirpref give me a good results for a wide range of
file operations and different filesystem capacities except one case:
those applications that create their entire directory structure first
and only later fill this structure with files.
My solution for such and similar cases is to limit a number of
directories which may be created one after another in the same cylinder
group without intervening file creations. For this purpose, I allocate
an array of counters at mount time. This array is linked to the superblock
fs->fs_contigdirs[cg]. Each time a directory is created the counter
increases and each time a file is created the counter decreases. A 60Gb
filesystem with 8mb/cg requires 10kb of memory for the counters array.
The maxcontigdirs is a maximum number of directories which may be created
without an intervening file creation. I found in my tests that the best
performance occurs when I restrict the number of directories in one cylinder
group such that all its files may be located in the same cylinder group.
There may be some deterioration in performance if all the file inodes
are in the same cylinder group as its containing directory, but their
data partially resides in a different cylinder group. The maxcontigdirs
value is calculated to try to prevent this condition. Since there is
no way to know how many files and directories will be allocated later
I added two optimization parameters in superblock/tunefs. They are:
int32_t fs_avgfilesize; /* expected average file size */
int32_t fs_avgfpdir; /* expected # of files per directory */
These parameters have reasonable defaults but may be tweeked for special
uses of a filesystem. They are only necessary in rare cases like better
tuning a filesystem being used to store a squid cache.
I have been using this algorithm for about 3 months. I have done
a lot of testing on filesystems with different capacities, average
filesize, average number of files per directory, and so on. I think
this algorithm has no negative impact on filesystem perfomance. It
works better than the default one in all cases. The new dirpref
will greatly improve untarring/removing/coping of big directories,
decrease load on cvs servers and much more. The new dirpref doesn't
speedup a compilation process, but also doesn't slow it down.
Obtained from: Grigoriy Orlov <gluk@ptci.ru>
the gating of system calls that cause modifications to the underlying
filesystem. The gating can be enabled by any filesystem that needs
to consistently suspend operations by adding the vop_stdgetwritemount
to their set of vnops. Once gating is enabled, the function
vfs_write_suspend stops all new write operations to a filesystem,
allows any filesystem modifying system calls already in progress
to complete, then sync's the filesystem to disk and returns. The
function vfs_write_resume allows the suspended write operations to
begin again. Gating is not added by default for all filesystems as
for SMP systems it adds two extra locks to such critical kernel
paths as the write system call. Thus, gating should only be added
as needed.
Details on the use and current status of snapshots in FFS can be
found in /sys/ufs/ffs/README.snapshot so for brevity and timelyness
is not included here. Unless and until you create a snapshot file,
these changes should have no effect on your system (famous last words).
<sys/bio.h>.
<sys/bio.h> is now a prerequisite for <sys/buf.h> but it shall
not be made a nested include according to bdes teachings on the
subject of nested includes.
Diskdrivers and similar stuff below specfs::strategy() should no
longer need to include <sys/buf.> unless they need caching of data.
Still a few bogus uses of struct buf to track down.
Repocopy by: peter
(name, value) pairs to be associated with inodes. This support is
used for ACLs, MAC labels, and Capabilities in the TrustedBSD
security extensions, which are currently under development.
In this implementation, attributes are backed to data vnodes in the
style of the quota support in FFS. Support for FFS extended
attributes may be enabled using the FFS_EXTATTR kernel option
(disabled by default). Userland utilities and man pages will be
committed in the next batch. VFS interfaces and man pages have
been in the repo since 4.0-RELEASE and are unchanged.
o ufs/ufs/extattr.h: UFS-specific extattr defines
o ufs/ufs/ufs_extattr.c: bulk of support routines
o ufs/{ufs,ffs,mfs}/*.[ch]: hooks and extattr.h includes
o contrib/softupdates/ffs_softdep.c: extattr.h includes
o conf/options, conf/files, i386/conf/LINT: added FFS_EXTATTR
o coda/coda_vfsops.c: XXX required extattr.h due to ufsmount.h
(This should not be the case, and will be fixed in a future commit)
Currently attributes are not supported in MFS. This will be fixed.
Reviewed by: adrian, bp, freebsd-fs, other unthanked souls
Obtained from: TrustedBSD Project
1) Fastpath deletions. When a file is being deleted, check to see if it
was so recently created that its inode has not yet been written to
disk. If so, the delete can proceed to immediately free the inode.
2) Background writes: No file or block allocations can be done while the
bitmap is being written to disk. To avoid these stalls, the bitmap is
copied to another buffer which is written thus leaving the original
available for futher allocations.
3) Link count tracking. Constantly track the difference in i_effnlink and
i_nlink so that inodes that have had no change other than i_effnlink
need not be written.
4) Identify buffers with rollback dependencies so that the buffer flushing
daemon can choose to skip over them.
basically do a on-the-fly defragmentation of the FFS filesystem, changing
file block allocations to make them contiguous. Thanks to Kirk McKusick
for providing hints on what needed to be done to get this working.
device drivers about sectors no longer in use.
Device-drivers receive the call through d_strategy, if they have
D_CANFREE in d_flags.
This allows flash based devices to erase the sectors and avoid
pointlessly carrying them around in compactions.
Reviewed by: Kirk Mckusick, bde
Sponsored by: M-Systems (www.m-sys.com)
"time" wasn't a atomic variable, so splfoo() protection were needed
around any access to it, unless you just wanted the seconds part.
Most uses of time.tv_sec now uses the new variable time_second instead.
gettime() changed to getmicrotime(0.
Remove a couple of unneeded splfoo() protections, the new getmicrotime()
is atomic, (until Bruce sets a breakpoint in it).
A couple of places needed random data, so use read_random() instead
of mucking about with time which isn't random.
Add a new nfs_curusec() function.
Mark a couple of bogosities involving the now disappeard time variable.
Update ffs_update() to avoid the weird "== &time" checks, by fixing the
one remaining call that passwd &time as args.
Change profiling in ncr.c to use ticks instead of time. Resolution is
the same.
Add new function "tvtohz()" to avoid the bogus "splfoo(), add time, call
hzto() which subtracts time" sequences.
Reviewed by: bde
are used in the `#ifdef notyet' case :-). This case is used except in
the `#if !defined (not_yes)' case :-|. This has something to do with
the `#ifdef notyet_block_reallocation_enabled' case in vfs_cluster.c :-(.
1. Remove VOP_UPDATE, it is (also) an UFS/{FFS,LFS,EXT2FS,MFS}
intereface function, and now lives in the ufsmount structure.
2. Remove VOP_SEEK, it was unused.
3. Add mode default vops:
VOP_ADVLOCK vop_einval
VOP_CLOSE vop_null
VOP_FSYNC vop_null
VOP_IOCTL vop_enotty
VOP_MMAP vop_einval
VOP_OPEN vop_null
VOP_PATHCONF vop_einval
VOP_READLINK vop_einval
VOP_REALLOCBLKS vop_eopnotsupp
And remove identical functionality from filesystems
4. Add vop_stdpathconf, which returns the canonical stuff. Use
it in the filesystems. (XXX: It's probably wrong that specfs
and fifofs sets this vop, shouldn't it come from the "host"
filesystem, for instance ufs or cd9660 ?)
5. Try to make system wide VOP functions have vop_* names.
6. Initialize the um_* vectors in LFS.
(Recompile your LKMS!!!)
1. Add new file "sys/kern/vfs_default.c" where default actions for
VOPs go. Implement proper defaults for ABORTOP, BWRITE, LEASE,
POLL, REVOKE and STRATEGY. Various stuff spread over the entire
tree belongs here.
2. Change VOP_BLKATOFF to a normal function in cd9660.
3. Kill VOP_BLKATOFF, VOP_TRUNCATE, VOP_VFREE, VOP_VALLOC. These
are private interface functions between UFS and the underlying
storage manager layer (FFS/LFS/MFS/EXT2FS). The functions now
live in struct ufsmount instead.
4. Remove a kludge of VOP_ functions in all filesystems, that did
nothing but obscure the simplicity and break the expandability.
If a filesystem doesn't implement VOP_FOO, it shouldn't have an
entry for it in its vnops table. The system will try to DTRT
if it is not implemented. There are still some cruft left, but
the bulk of it is done.
5. Fix another VCALL in vfs_cache.c (thanks Bruce!)
1. ffs_alloc() actually allowed writing one block less one frag (normally
7 frags or 7/8 blocks) beyond the limit.
2. freebufspace() gives the free space in frags, but `size' is in bytes,
so the change results in approximately `size' fragments too many being
reserved.
3. ffs_realloccg() has the same bug but wasn't changed.
PR: 3398
Submitted by: bde
Eyeballed by: phk
code. According to the crash dump, bpref is set to 445
and cgp->cg_nclusterblks is 444. Hence in the for loop,
the test fails immediately but the following failure check
(got == cgp->cg_nclusterblks) doesn't trigger because got >
cgp->cg_nclusterblks. This wreaks havoc in the code after that.
Fix: Move one source bit to the left :-)
Noticed by: Mike Hibler <mike@fast.cs.utah.edu>
Submitted by: Kirk McKusick <mckusick@McKusick.COM>
automatically have random generation numbers. The kenel way of handling those
also changed. Further it is advised to run fsirand on all your nfs exported
filesystems. the code is mostly copied from OpenBSD, with the randomization
chanegd to use /dev/urandom
Reviewed by: Garrett
Obtained from: OpenBSD
form `tv = time'. Use a new function gettime(). The current version
just forces atomicicity without fixing precision or efficiency bugs.
Simplified some related valid accesses by using the central function.
to not return without setting a return value when it
can't read a block error or detects a bad cylinder group,
since the caller is expecting a return value.
It will now panic at this point, since the thing
to do in this case would be to return a "bad block"
status to the caller, and the caller will panic
anyways when that happens.
Also updated to panic strings in this routine to read
"ffs_checkblk: ..." instead of "checkblk: ...".
changes, so don't expect to be able to run the kernel as-is (very well)
without the appropriate Lite/2 userland changes.
The system boots and can mount UFS filesystems.
Untested: ext2fs, msdosfs, NFS
Known problems: Incorrect Berkeley ID strings in some files.
Mount_std mounts will not work until the getfsent
library routine is changed.
Reviewed by: various people
Submitted by: Jeffery Hsu <hsu@freebsd.org>
