[OpenAFS-devel] PATCH: limit afsd auto-tuning to 500000 files

[Niklas Edmundsson]

On Tue, 23 Aug 2005, Jeffrey Hutzelman wrote:

>> In message <475D0B15E1CC68BBF87A7B2D@sirius.fac.cs.cmu.edu>,Jeffrey
>> Hutzelman writes:
>>> 67%-full assumption is incorrect.  Perhaps we were too conservative in
>>> bumping the average filesize assumption from 10K to 32K, and it should
>>> really be bigger.
>> 
>> i dont believe that there should be linear relationsip between the
>> average filesize and the cache size.  when a user increases the disk
>> cache, the more kernel memory is not available.
>
> Well, right now we use two numbers.  One is a constant; the other is a 
> function of the chunk size.  It sounds like you're arguing for eliminating 
> the constant, or at least limiting its effect as the cache size grows very 
> large.  Fine, but without data, how do we decide where to draw the line? With 
> a small cache and a large chunk size, we need that constant to ensure we have 
> enough files.

There was a suggestion not that long ago to use sqrt() to calculate 
the average filesize and thus obtain the limiting effect. Arbitrary, 
yes, but it did provide nice values given the assumption that large 
caches means larger average filesizes. It's certainly better than the 
hard coded average filesize, and should be enough to calm the 
situation down until a thorough investigation has been made.

I also suspect that the main reason that people want to limit the 
number of files is that certain platforms (32bit x86 Linux comes to 
mind) has a limited amount of kernel memory, and thus fails hard if 
the user creates a cache which has too many files.

/Nikke
-- 
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
  Niklas Edmundsson, Admin @ {acc,hpc2n}.umu.se     |    nikke@hpc2n.umu.se
---------------------------------------------------------------------------
  It's curtains for Windows!
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=