[OpenAFS] cache performance

[Phil.Moore@morganstanley.com]

>>>>> "Nickolai" == Nickolai Zeldovich <kolya@MIT.EDU> writes:

>> What do you think about using the mount options "noatime,data=writeback" 
>> when using an ext3 filesystem for the AFS cache on linux systems.
>> I think this could increase the performance on a heavily used cache.

Nickolai> The "noatime" option will not make any difference -- the kernel module
Nickolai> already sets the NOATIME bit on all the cache file inodes.  I don't know
Nickolai> for sure what the "data=writeback" option does, but guessing that it makes
Nickolai> the buffer cache a write-back one, sure, it may increase your performance.
Nickolai> Though I had the impression that Linux already did writeback caching..  Or
Nickolai> maybe that was only for ext2, and ext3 enforces more traditional semantics?

NOATIME!?  Uh, forgive my ignorance, but I take it this means that
access times are no longer maintained on the V files in the cache?

I understand why this would be the case, as an obvious performance
improvement, but I have code that analyzes the AFS cache and
summarizes the usage by volume and cell, showing the size and
*lastaccess* time of the volume in the cache.

Experimentally, simply stating the V files and taking the most recent
last access timestampt for all of the V files that comprise a specific
volume did in fact produce results that were contistent with my
expectations.

The reason we have this code is that we analyze the contents of ALL of
our clients caches (yes, I'm not joking, we really do), and procude
enterprise wide reports on who is accessing what volumes.  This has
proven very useful.

Now, what you are saying suggests that we can't contoinue to make this
assumption about the access times, and thus, we'll have to rethink our
cache analysis code.