Re: [RFC] Improved inode number allocation for HTree

Re: [RFC] Improved inode number allocation for HTree

[Pascal Schmidt]

On Tue, 11 Mar 2003 20:40:14 +0100, you wrote in linux.kernel:

> Ok, so "rm" works.  Then you have things like "mv *.c /usr/src" to worry
> about.  Lock for traversal, get stuck unable to work on the files.

In both cases, the shell does the traversal and passes a complete list
of files to rm or mv... so rm and mv themselves don't need to do any
directory traversal.

-- 
Ciao,
Pascal

[Bug 417] New: htree much slower than regular ext3

[Martin J. Bligh]

http://bugme.osdl.org/show_bug.cgi?id=417

           Summary: htree much slower than regular ext3
    Kernel Version: 2.5.63-bk3
            Status: NEW
          Severity: normal
             Owner: akpm@digeo.com
         Submitter: bwindle-kbt@fint.org


Distribution: Debian Testing

Hardware Environment: x86, 256mb RAM, PIIX4, Maxtor 91728D8 ATA DISK drive

Software Environment: 
Linux razor 2.5.63bk3 #25 Thu Feb 27 10:13:35 EST 2003 i686 Pentium II 
(Klamath) GenuineIntel GNU/Linux

Gnu C                  2.95.4
Gnu make               3.79.1
util-linux             2.11n
mount                  2.11n
module-init-tools      implemented
e2fsprogs              1.30-WIP
reiserfsprogs          3.6.3
Linux C Library        2.2.5
Dynamic linker (ldd)   2.2.5
Procps                 2.0.7
Net-tools              1.60
Console-tools          0.2.3
Sh-utils               4.5.2

Problem Description:
I created a directory ("test") with 32000 (ok, 31998) directories in it,
and  put a file called 'foo' in each of them (for i in `ls`; do cd $i &&
touch bar  && cd .. ; done). Then I took that ext3 partion, umounted it,
did a 'tune2fs -O  dir_index', then 'fsck -fD', and remounted. I then did a
'time du -hs' on the  test directory, and here are the results.

ext3+htree:     
bwindle@razor:/giant/inodes$ time du -hs
126M    .

real    7m21.756s
user    0m2.021s
sys     0m22.190s

I then unmounted, tune2fs -O ^dir_index, e2fsck -fD /dev/hdb1, remounted,
and  did another du -hs on the test directory. It took 1 minute, 48 seconds.

bwindle@razor:/giant/test$ time du -hs
126M    .

real    1m48.760s
user    0m1.986s
sys     0m21.563s


I thought htree was supposed to speed up access with large numbers of 
directories?

[RFC] Improved inode number allocation for HTree

[Daniel Phillips]

Summary of the problem:

HTree's random relationship between directory entry order and inode order has 
been observed to cause increased cache pressure, affecting performance under 
certain conditions.  Most of this pressure is due to inode table usage, which 
is normally several times larger than the directory itself.  Even when the 
inode table usage is much less than available cache, it may still exceed the 
size of the journal file (8 MB by default).  Though the journal only becomes 
involved when blocks are modified, unfortunately, because of atime updates, 
this includes all directory operations.  We could suggest to users that they 
should disable atime updating if they care about performance, but we ought to 
be able to do better than that.

To reduce this cache pressure, what we need to do is make the directory entry 
order correspond more closely to the inode number order.  It doesn't have to 
correspond perfectly in order to show substantial improvement, as Tomas 
demonstrated with his recent getdents sorting patch.

Analysis:

To understand why sorting relatively small sections of the directory entries 
reduces the cache pressure, we can think about the lifetime of an inode table 
block over the course of a directory traversal.  This traversal lifetime 
extends from when it is first accessed (always a write as noted above) to 
when it is last accessed, or the directory traversal finishes.  Each inode 
table block contains multiple inodes, so if those inodes are accessed 
randomly, the traversal lifetime of every inode table block will tend to be 
from nearly the beginning of the traversal to nearly the end, and so the load 
on the cache will be nearly the same as the number of inode table blocks.  If 
this load is greater than available cache, various blocks will have to be 
written out and reloaded later for further processing.  We can call the 
period between loading and evicting an inode table block, its "cache period".

The number of times a block must be written/reloaded depends not only on the 
number of inode table blocks covered by the directory and the availablity of 
cache memory, but on the probable number of the inodes within the block that 
will be written during each of the block's cache period.  By sorting the 
directory chunkwise, the probable number of inodes processed in one cache 
period increases, reducing the number of cache periods, in turn reducing disk 
seeking and transfer overhead.

To apply the above model to the journal, just substitute "journal period" for 
"cache period", where journal period is defined as the time from when a dirty 
block first enters the journal to when it is written out to disk. 

We can see that local sorting or directory chunks has hardly any effect on 
the traversal lifetime of a block; it only reduces the number of cache 
periods.  The size of the chunk of directory that we sort locally presumably 
stays constant, so the probable number of inodes on each inode table block 
processed in each cache period approaches one as the directory size 
increases.  At some point, there is no longer any benefit from sorting, andwe 
arrive at that point at rather realistic directory sizes.

We can always do the job perfectly by sorting the whole directory.  This 
would require changes in every getdents-using application that cares about 
performance, to either make the getdents window as big as the directory or to 
do the sort in user space.  A spinoff benefit of doing the full sort is that 
it becomes trivial to detect and eliminate duplicate entries.  Higher memory 
load is a drawback, as is the work required to update user space programs.  
There will also be noticable latency introduced by the sort, which didn't 
appear before.  If sorting is to be done in user space then we impose the 
requirement on all filesystems - even those that do not use inode numbers 
internally - that they supply an inode key for sorting, and that the supplied 
sorting key be meaningful in terms of cache locality.

Proposed Solution:

What if we could maintain the inodes in the inode table in the same order as 
the directory entries, permanently, so that no sorting is required?  If we 
could do this, the traversal lifetime of each inode table block would be the 
same as the number of inodes it contains, and the cache load due to the inode 
table would be exactly one block.  That is, the directory traversal would 
process each inode table block completely before moving on to the next one. 
Though the current Ext3 filesystem structure does not support this directly, 
we can approximate the effect quite closely within the existing framework.

Due to Ted's recent work, we are already traversing the directory in hash 
order, as opposed to physical allocation order, because of the particular 
approach adopted to address the telldir/seekdir problem.  This is relatively 
inexpensive.  So now, what we would like to do is make the inode number of 
each directory entry increase monotonically, corresponding to monotonically 
increasing hash values.  We also want to allocate those inode numbers densely 
in the inode table, to minimize seeking.  If we know beforehand how many 
entries the directory will have, this is easy: when we create an entry, we 
multiply its hash value by the total number of entries divided by the total 
hash range (2**31), add in a base target, and use that as the goal to 
allocate a new inode.  Due to the randomness of the hash value, we will 
sometimes be forced to allocate an inode number out of order, but hopefully 
there won't be too many of those, and they won't be very far out of order.  
We may also end up with some gaps in the inode table.  In fact, we may 
intentionally leave some slack in order to reduce the number of out-of-order 
allocations.  These problems are not serious.

The big problem is that we don't know the number of directory entries ahead 
of time.  To get around this, we can use the current size of the directory, 
rounded up to a power of two as the size estimate.  Each time the directory
increases to a new binary size we establish a new region in the inode table, 
which will map to the to-be-created half of directory entries, roughly in 
order.

Now, consider what happens to a group of directory entries created when the 
directory was so small that it mapped to a single inode table block.  After a 
series of leaf splits, these former neighbors will end up far apart in the 
hash order, but their inode numbers will still be ordered monotonically over 
a traversal.  The traversal lifetime of that group of entries will be nearly 
the entire directory traversal, but there is only one such block.

When the directory became larger, some newly created entries were mapped to 
two new inode table blocks, the traversal lifetimes of which are roughly half 
the entire traversal.  There are only two such blocks, and their lifetimes do 
not overlap, so these two blocks only generate one block worth of cache load.

So it goes, with each new set of entries of size 2**i generating only one 
additional block of cache load.  The final cache load is O log2(N), a 
pleasingly small number.

A very nice property of this inode allocation strategy is that the ordering 
will tend to be reinforced as a directory ages, as opposed to slowly decaying 
towards randomness as happens with linear directories.

So this approach is attractive, especially as it can be implemented in a 
small number of lines of code.  The main outstanding issue I have not 
addressed is how to arrange things so that the inode allocation patterns of 
neighbouring directories don't interfere with each other too much.  I expect 
that each time the directory grows to a new power of two size, we need to 
scan the inode allocation map for a relatively unpopulated region of that 
size and store the region's base in the directory header.  Would we need to 
store the entire history of allocation bases?  Probably not.  This aspect 
requires further consideration.

Regards,

Daniel

Re: [RFC] Improved inode number allocation for HTree

[Hans Reiser]

Let's make noatime the default for VFS.

Daniel Phillips wrote:

>Hi Hans,
>
>On Tue 11 Mar 03 00:17, Hans Reiser wrote:
>  
>
>>What do you think of creating a new telldir/seekdir which uses filenames
>>instead of ints to convey position?
>>    
>>
>
>What do you do if that file gets deleted in the middle of a traversal?
>
So what?  It still tells you where to restart.  Strictly speaking, you 
would want to allow the filesystem to choose what it wants to return to 
indicate position.  For htree and reiserfs this would be a filename or 
its hash, for ext2 without htree this would be a byte offset.

>
>If I were able to design Unix over again, I'd state that if you don't lock a 
>directory before traversing it then it's your own fault if somebody changes 
>it under you, and I would have provided an interface to inform you about your 
>bad luck.  Strictly wishful thinking.  (There, it feels better now.)
>
We are designing Linux.  You know, Microsoft and Steve Jobs continue to 
design.  We should too.

Needless change should be avoided.  Failure to change when something is 
known to be broken leads to being inferior to those who do change.

Let's design something that does it right.  Or else SCO will be right in 
saying that Linux is just a Unix ripoff by people who couldn't do 
anything unless Unix had been written to tell them how to do it right.

-- 
Hans

Re: [RFC] Improved inode number allocation for HTree

[Helge Hafting]

Hans Reiser wrote:
> Let's make noatime the default for VFS.
> 
> Daniel Phillips wrote:
[...]
>> If I were able to design Unix over again, I'd state that if you don't 
>> lock a directory before traversing it then it's your own fault if 
>> somebody changes it under you, and I would have provided an interface 
>> to inform you about your bad luck.  Strictly wishful thinking.  
>> (There, it feels better now.)

I'm happy nobody _can_ lock a directory like that.  Think of it - unable
to create or delete files while some slow-moving program is traversing
the directory?  Ouch.  Plenty of options for DOS attacks too.
And how to do "rm *.bak" if rm locks the dir for traversal?

Helge Hafting

Re: [RFC] Improved inode number allocation for HTree

[Daniel Phillips]

On Tue 11 Mar 03 14:00, Helge Hafting wrote:
> Hans Reiser wrote:
> > Let's make noatime the default for VFS.
> >
> > Daniel Phillips wrote:
>
> [...]
>
> >> If I were able to design Unix over again, I'd state that if you don't
> >> lock a directory before traversing it then it's your own fault if
> >> somebody changes it under you, and I would have provided an interface
> >> to inform you about your bad luck.  Strictly wishful thinking.
> >> (There, it feels better now.)
>
> I'm happy nobody _can_ lock a directory like that.  Think of it - unable
> to create or delete files while some slow-moving program is traversing
> the directory?  Ouch.  Plenty of options for DOS attacks too.
> And how to do "rm *.bak" if rm locks the dir for traversal?

<wishful thinking>
Now that you mention it, just locking out create and rename during directory 
traversal would eliminate the pain.  Delete is easy enough to handle during 
traversal.  For a B-Tree, coalescing could simply be deferred until the 
traversal is finished, so reading the directory in physical storage order 
would be fine.  Way, way cleaner than what we have to do now.
</wishful thinking>

Regards,

Daniel

Re: [RFC] Improved inode number allocation for HTree

[Andreas Dilger]

On Mar 11, 2003  14:41 +0100, Daniel Phillips wrote:
> On Tue 11 Mar 03 14:00, Helge Hafting wrote:
> > I'm happy nobody _can_ lock a directory like that.  Think of it - unable
> > to create or delete files while some slow-moving program is traversing
> > the directory?  Ouch.  Plenty of options for DOS attacks too.
> > And how to do "rm *.bak" if rm locks the dir for traversal?
> 
> <wishful thinking>
> Now that you mention it, just locking out create and rename during directory 
> traversal would eliminate the pain.  Delete is easy enough to handle during 
> traversal.  For a B-Tree, coalescing could simply be deferred until the 
> traversal is finished, so reading the directory in physical storage order 
> would be fine.  Way, way cleaner than what we have to do now.
> </wishful thinking>

The other problem, of course, is that this would essentially mean that a
user-space application has a lock that prevents other processes (even the
kernel) from doing anything.  At best we have to worry about an application
going sour, and at worst it is a DOS hole as Helge says.

How about something like:

	death:
		for each directory
			while readdir(directory, large_buffer)
				if (directory)
					run death(directory)
			readdir(directory, small_buffer)

		sleep forever

The first getdents walks recursively down to the leaves, and then does a
single getdents with enough space to hold a single dirent, locking the
directory, but then never progressing.  It also does the single-dirent
getdents on the way up the tree.  Now, everything is locked out of the
entire directory tree.

Cheers, Andreas
--
Andreas Dilger
http://sourceforge.net/projects/ext2resize/
http://www-mddsp.enel.ucalgary.ca/People/adilger/

Re: [RFC] Improved inode number allocation for HTree

[Helge Hafting]

On Tue, Mar 11, 2003 at 02:41:06PM +0100, Daniel Phillips wrote:
> On Tue 11 Mar 03 14:00, Helge Hafting wrote:
> > I'm happy nobody _can_ lock a directory like that.  Think of it - unable
> > to create or delete files while some slow-moving program is traversing
> > the directory?  Ouch.  Plenty of options for DOS attacks too.
> > And how to do "rm *.bak" if rm locks the dir for traversal?
> 
> <wishful thinking>
> Now that you mention it, just locking out create and rename during directory 
> traversal would eliminate the pain.  Delete is easy enough to handle during 
> traversal.  For a B-Tree, coalescing could simply be deferred until the 
> traversal is finished, so reading the directory in physical storage order 
> would be fine.  Way, way cleaner than what we have to do now.
> </wishful thinking>

Ok, so "rm" works.  Then you have things like "mv *.c /usr/src" to worry
about.  Lock for traversal, get stuck unable to work on the files.

A "dream" solution for this might involve something like:
Directory is traversed in some well defined order (defined by the fs)
Removing a file (or moving it out of the directory) leaves
a "empty" directory entry that isn't reclaimed until no more
traversals is in progress in that directory.

New files can be created, and will be created further out than
any traversal in progress so nothing will be missed.

This approach don't lock anything out, but I guess someone evil still
could cause trouble by keeping a traversal going forever, creating one dummy
file and deleting one whenever it makes progress.  The directory would
get big, filled up with placeholders until some ulimit kicks in.

Helge Hafting

Re: [RFC] Improved inode number allocation for HTree

[Daniel Phillips]

On Tue 11 Mar 03 20:39, Helge Hafting wrote:
> On Tue, Mar 11, 2003 at 02:41:06PM +0100, Daniel Phillips wrote:
> > <wishful thinking>
> > Now that you mention it, just locking out create and rename during
> > directory traversal would eliminate the pain.  Delete is easy enough to
> > handle during traversal.  For a B-Tree, coalescing could simply be
> > deferred until the traversal is finished, so reading the directory in
> > physical storage order would be fine.  Way, way cleaner than what we have
> > to do now.
> > </wishful thinking>
>
> Ok, so "rm" works.  Then you have things like "mv *.c /usr/src" to worry
> about.

That's equivalent to ls, the shell expands it before doing any moving.  You 
can construct something more interesting with ls | xargs <something nasty>
into the same directory.  Since the user is trying to shoot their foot off, 
let the lock be recursive, and let them do that.

> ...someone evil still
> could cause trouble by keeping a traversal going forever, creating one
> dummy file and deleting one whenever it makes progress.  The directory
> would get big, filled up with placeholders until some ulimit kicks in.
>
> Helge Hafting

It's not clear that's any more evil than things they can do already, eg,

   seq 1 1000000 | xargs -l1 cp -a /usr

Welllll, this is all idle speculation anyway.  Nobody's going to fix the 
flaw this week, if it's even possible (I suspect it is).

Regards,

Daniel

[RFC] Improved inode number allocation for HTree

[Daniel Phillips]

Summary of the problem:

HTree's random relationship between directory entry order and inode order has 
been observed to cause increased cache pressure, affecting performance under 
certain conditions.  Most of this pressure is due to inode table usage, which 
is normally several times larger than the directory itself.  Even when the 
inode table usage is much less than available cache, it may still exceed the 
size of the journal file (8 MB by default).  Though the journal only becomes 
involved when blocks are modified, unfortunately, because of atime updates, 
this includes all directory operations.  We could suggest to users that they 
should disable atime updating if they care about performance, but we ought to 
be able to do better than that.

To reduce this cache pressure, what we need to do is make the directory entry 
order correspond more closely to the inode number order.  It doesn't have to 
correspond perfectly in order to show substantial improvement, as Tomas 
demonstrated with his recent getdents sorting patch.

Analysis:

To understand why sorting relatively small sections of the directory entries 
reduces the cache pressure, we can think about the lifetime of an inode table 
block over the course of a directory traversal.  This traversal lifetime 
extends from when it is first accessed (always a write as noted above) to 
when it is last accessed, or the directory traversal finishes.  Each inode 
table block contains multiple inodes, so if those inodes are accessed 
randomly, the traversal lifetime of every inode table block will tend to be 
from nearly the beginning of the traversal to nearly the end, and so the load 
on the cache will be nearly the same as the number of inode table blocks.  If 
this load is greater than available cache, various blocks will have to be 
written out and reloaded later for further processing.  We can call the 
period between loading and evicting an inode table block, its "cache period".

The number of times a block must be written/reloaded depends not only on the 
number of inode table blocks covered by the directory and the availablity of 
cache memory, but on the probable number of the inodes within the block that 
will be written during each of the block's cache period.  By sorting the 
directory chunkwise, the probable number of inodes processed in one cache 
period increases, reducing the number of cache periods, in turn reducing disk 
seeking and transfer overhead.

To apply the above model to the journal, just substitute "journal period" for 
"cache period", where journal period is defined as the time from when a dirty 
block first enters the journal to when it is written out to disk. 

We can see that local sorting or directory chunks has hardly any effect on 
the traversal lifetime of a block; it only reduces the number of cache 
periods.  The size of the chunk of directory that we sort locally presumably 
stays constant, so the probable number of inodes on each inode table block 
processed in each cache period approaches one as the directory size 
increases.  At some point, there is no longer any benefit from sorting, andwe 
arrive at that point at rather realistic directory sizes.

We can always do the job perfectly by sorting the whole directory.  This 
would require changes in every getdents-using application that cares about 
performance, to either make the getdents window as big as the directory or to 
do the sort in user space.  A spinoff benefit of doing the full sort is that 
it becomes trivial to detect and eliminate duplicate entries.  Higher memory 
load is a drawback, as is the work required to update user space programs.  
There will also be noticable latency introduced by the sort, which didn't 
appear before.  If sorting is to be done in user space then we impose the 
requirement on all filesystems - even those that do not use inode numbers 
internally - that they supply an inode key for sorting, and that the supplied 
sorting key be meaningful in terms of cache locality.

Proposed Solution:

What if we could maintain the inodes in the inode table in the same order as 
the directory entries, permanently, so that no sorting is required?  If we 
could do this, the traversal lifetime of each inode table block would be the 
same as the number of inodes it contains, and the cache load due to the inode 
table would be exactly one block.  That is, the directory traversal would 
process each inode table block completely before moving on to the next one. 
Though the current Ext3 filesystem structure does not support this directly, 
we can approximate the effect quite closely within the existing framework.

Due to Ted's recent work, we are already traversing the directory in hash 
order, as opposed to physical allocation order, because of the particular 
approach adopted to address the telldir/seekdir problem.  This is relatively 
inexpensive.  So now, what we would like to do is make the inode number of 
each directory entry increase monotonically, corresponding to monotonically 
increasing hash values.  We also want to allocate those inode numbers densely 
in the inode table, to minimize seeking.  If we know beforehand how many 
entries the directory will have, this is easy: when we create an entry, we 
multiply its hash value by the total number of entries divided by the total 
hash range (2**31), add in a base target, and use that as the goal to 
allocate a new inode.  Due to the randomness of the hash value, we will 
sometimes be forced to allocate an inode number out of order, but hopefully 
there won't be too many of those, and they won't be very far out of order.  
We may also end up with some gaps in the inode table.  In fact, we may 
intentionally leave some slack in order to reduce the number of out-of-order 
allocations.  These problems are not serious.

The big problem is that we don't know the number of directory entries ahead 
of time.  To get around this, we can use the current size of the directory, 
rounded up to a power of two as the size estimate.  Each time the directory
increases to a new binary size we establish a new region in the inode table, 
which will map to the to-be-created half of directory entries, roughly in 
order.

Now, consider what happens to a group of directory entries created when the 
directory was so small that it mapped to a single inode table block.  After a 
series of leaf splits, these former neighbors will end up far apart in the 
hash order, but their inode numbers will still be ordered monotonically over 
a traversal.  The traversal lifetime of that group of entries will be nearly 
the entire directory traversal, but there is only one such block.

When the directory became larger, some newly created entries were mapped to 
two new inode table blocks, the traversal lifetimes of which are roughly half 
the entire traversal.  There are only two such blocks, and their lifetimes do 
not overlap, so these two blocks only generate one block worth of cache load.

So it goes, with each new set of entries of size 2**i generating only one 
additional block of cache load.  The final cache load is O log2(N), a 
pleasingly small number.

A very nice property of this inode allocation strategy is that the ordering 
will tend to be reinforced as a directory ages, as opposed to slowly decaying 
towards randomness as happens with linear directories.

So this approach is attractive, especially as it can be implemented in a 
small number of lines of code.  The main outstanding issue I have not 
addressed is how to arrange things so that the inode allocation patterns of 
neighbouring directories don't interfere with each other too much.  I expect 
that each time the directory grows to a new power of two size, we need to 
scan the inode allocation map for a relatively unpopulated region of that 
size and store the region's base in the directory header.  Would we need to 
store the entire history of allocation bases?  Probably not.  This aspect 
requires further consideration.

Regards,

Daniel

Re: [RFC] Improved inode number allocation for HTree

[John Bradford]

> Though the journal only becomes involved when blocks are modified,
> unfortunately, because of atime updates, this includes all directory
> operations.  We could suggest to users that they should disable
> atime updating if they care about performance, but we ought to be
> able to do better than that.

On a separate note, since atime updates are not usually very important
anyway, why not have an option to cache atime updates for a long time,
or until either a write occurs anyway.  Holding a large number of
atime updates in a write cache is generally not going to be a major
issue - the worst case if a partition isn't cleanly unmounted is that
some atimes will be wrong.

John.

Re: [RFC] Improved inode number allocation for HTree

[Andreas Schwab]

John Bradford <john@grabjohn.com> writes:

|> > Though the journal only becomes involved when blocks are modified,
|> > unfortunately, because of atime updates, this includes all directory
|> > operations.  We could suggest to users that they should disable
|> > atime updating if they care about performance, but we ought to be
|> > able to do better than that.
|> 
|> On a separate note, since atime updates are not usually very important
|> anyway, why not have an option to cache atime updates for a long time,
|> or until either a write occurs anyway.

mount -o noatime

Andreas.

-- 
Andreas Schwab, SuSE Labs, schwab@suse.de
SuSE Linux AG, Deutschherrnstr. 15-19, D-90429 Nürnberg
Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

Re: [RFC] Improved inode number allocation for HTree

[Daniel Phillips]

On Mon 10 Mar 03 22:04, John Bradford wrote:
> > Though the journal only becomes involved when blocks are modified,
> > unfortunately, because of atime updates, this includes all directory
> > operations.  We could suggest to users that they should disable
> > atime updating if they care about performance, but we ought to be
> > able to do better than that.
>
> On a separate note, since atime updates are not usually very important
> anyway, why not have an option to cache atime updates for a long time,
> or until either a write occurs anyway.  Holding a large number of
> atime updates in a write cache is generally not going to be a major
> issue - the worst case if a partition isn't cleanly unmounted is that
> some atimes will be wrong.

It sounds practical.  Why stop there?  Since Ted is seriously considering 
making a batch of incompatible extensions to the on-disk format anyway, how 
about adding an atime table to each block group, four bytes per inode.  Even 
without lazy updating, it would cut down the dirty  blocks generated by r/o 
operations a lot.  If actual atime is the latest of the atime table value and 
the inode atime value[1], then inode write operations won't generate extra 
traffic.  You will only get new traffic when somebody wants the real atime.  
I'd put this under the category of "things to add to Ted's long list of fun 
new ideas for Ext3/4".

[1] How to handle wrapping is left as an exercise for the interested reader.

Regards,

Daniel