FYI - see comment from Timur below.
-----Original Message-----
From: Timur Perelmutov [mailto:[log in to unmask]]
Sent: 02 October 2006 18:36
To: Synge, Owen
Cc: Jensen, J (Jens); Jon A. Bakken; Patrick Fuhrmann
Subject: Re: Next version of GLUE schema writeup
[log in to unmask] wrote:
> On Mon, 02 Oct 2006 15:53:02 +0100
> "Jensen, J (Jens)" <[log in to unmask]> wrote:
>
>
>> Maarten Litmaath wrote:
>>
>>> Jensen, J (Jens) wrote:
>>>
>>>
>>>> Hi Jean-Philippe.
>>>>
>>>> OK, fair comment - Owen and I discussed last week whether it was
>>>> getting too complicated (preferring to keep it simple).
>>>>
>>>> GridPP *does* use the accounting information today:
>>>> http://www.gridpp.ac.uk/storage/status/gridppDiscStatus.html
>>>>
>>> I have serious doubts about some of those numbers,
>>> exactly because of the sad state of affairs in the info system.
>>>
>> Quite. We are working on improving them. They are probably
>> mostly accurate; it's just doing accounting per VO which is
>> tricky.
>>
>
> The double accounting of DPM and D-Cache is partly to blame here.
>
>
>> We also know that some of this accounting is quite expensive
>> - for dCache, the space used is measured by running 'du' on
>> the pnfs filesystem.
>>
>
> This is true, it can be expensive. Running 'du' on the pnfs file system
> is how some UK sites do accounting with D-cache. D-cache also provides a
> working accounting publisher, but it requires that sites bind pools to
> pool groups with the VO name. This gives accurate reporting on how much
> space is used in a poolGroup, including replicas. The du method is slow,
> and has its own inaccuracies particularly replicas.
>
>
Running du on PNFS should not be used at all. First of all some files in
pnfs might be stored on multiple disks, while others are stored on the
tape only, so it will give no useful info to the client. Second it will
be extremely expensive for PNFS server and will slow down PNFS access
for other users.
> Regards
>
> Owen
>
|