[Pvfs2-developers] Fwd: [Pvfs2-users] metadata

[Sam Lang]

I'm looking at the statfs state machine on the client based on this  
users's questions.  It looks like there's a bug in that sm, because  
we just do a statfs to each of the servers (both IO and metadata).   
If the metadata and IO nodes are split, and the storage on the  
metadata nodes has a smaller capacity (i could imagine cases where it  
was much smaller), then the df results will be pretty inaccurate  
(since its the smallest capacity server * number of servers).

-sam

Begin forwarded message:

> From: Sam Lang <slang at mcs.anl.gov>
> Date: February 2, 2007 4:28:38 PM CST
> To: Sam Lang <slang at mcs.anl.gov>
> Cc: Steve <steve at bov.nu>, <pvfs2-users at beowulf-underground.org>
> Subject: Re: [Pvfs2-users] metadata
>
>
> On Feb 2, 2007, at 4:15 PM, Sam Lang wrote:
>
>>
>> On Feb 2, 2007, at 3:28 PM, Steve wrote:
>>
>>> Hi,
>>>
>>> Is there any rough rule of thumb guide for the size of the  
>>> metadata for the
>>> size of the store ?
>>
>> Hi Steve,
>>
>> The size of the metadata as a percent of the size of the storage  
>> space depends on the sizes of the files.  I'm doing a back of the  
>> envelope calculation, but I think that in general, each file takes  
>> up somewhere around 650 bytes to store its metadata, while  
>> directories are a little less (maybe 500 bytes).  Remember though  
>> that the metadata size changes for different configurations (more  
>> servers, etc.).  There's also overhead related to the storage  
>> space itself, which is about (100K * numservers).  If you know how  
>> many files/directories you have in your volume, you should be able  
>> to get a rough estimate of your metadata size using those numbers  
>> above.
>
> I should clarify though, that the filesystem size (capacity and  
> available) reported by df for PVFS volumes is just the sum of the  
> sizes on each of the IO servers (capacity is actually the min of  
> all servers * numservers).  This means that what df reports in  
> terms of usage can be inaccurate if the PVFS storage on the IO  
> nodes is shared with other files.
>
> -sam
>
>
>> -sam
>>
>>
>>>
>>> Steve
>>>
>>> _______________________________________________
>>> Pvfs2-users mailing list
>>> Pvfs2-users at beowulf-underground.org
>>> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
>>>
>>
>