On 06/04/11 18:00, Peter Grandi wrote:
> On 31/03/11 14:15, Peter Grandi wrote:
>
> [ ... ]
>
>> This threads on the XFS mailing list about 3ware and WD green issues:
>>
>> http://oss.sgi.com/archives/xfs/2011-02/msg00226.html
>
> Just noticed a generic Adaptec HA comment in the Lustre mailing list:
>
> http://lists.lustre.org/pipermail/lustre-discuss/2011-April/015428.html
Another interesting issue with power saving mode (load/unload
cycles have been previously an issue with laptop drives):
http://forums.whirlpool.net.au/archive/1367904
«Because these disks are being used in Linux, it basically
causes unending park/unparks of the head, due to the way Linux's
I/O works. It is my understanding that Linux reads data into a
buffer every 20 seconds, so therefore the data was being read,
then 8 seconds later the head was being parked, then another 12
seconds later more data was read.»
This is not likely to apply to very active arrays like those
used for LHC data processing, but it is something I'd watch out
for anyhow. BTW the «Linux reads data into a buffer every 20
seconds» may be confusedly referring to the usually extremely
inappropriate Linux kernel default flusher settings that allow
very large amounts of unwritten pages to accumulate in memory,
and then flush them all at once which can have unhappy
consequences. In general, but in particular to work around some
suboptimal behaviour by some 3ware HAs I have set:
vm/dirty_ratio =50
vm/dirty_background_ratio =2
vm/dirty_expire_centisecs =200
vm/dirty_writeback_centisecs =100
In some patched and/or recent kernel versions one can set the
dirty page hi water mark as a number of bytes/pages instead of
as a percentage of memory, which is far more appropriate.
|