Tux3 filesystem interest ()

Re: forward logging and NVRAM

Hi Timothy,

On Sunday 27 July 2008 14:10, timothy norman huber wrote:
> Daniel,
> 
> Several weeks ago, during a discussion over coffee, you mentioned a  
> rather intriguing optimization- by writing the start of the forward  
> log chain to NVRAM rather than disk, you avoid a disk seek and write  
> transfer for each forward log chain.  i believe there was another  
> motivation that had to do with avoiding traversing the entire volume  
> in case of a crash.  Was this design feature to overcome lame fsck  
> performance?

The same motivation, actually.  Writing the start of a forward log
chain to nvram instead of to some known location on disk means that I
do not have to worry about doing unnatural things to optimize such
loads as O_SYNC writing, where each transaction has to complete
before the next one can begin, which means that the forward log chain
never gets more than one element log, requiring a seek to the known
location for every transaction.  Not worse than journalling to be
sure, and usually better, but not as fast as avoiding a seek and disk
write per transaction entirely.

So if somebody could give me 16 bytes of NVRAM or so per volume I
would be most pleased and could use that to generate some nice O_SYNC
benchmarks 

By the way, I wonder why your posts do not show up in the Mailman
(Pipermail) archive, while mine and Matt's do?

(Continue reading)

headers

Maciej Żenczykowski | 29 Jul 20:06

Re: forward logging and NVRAM

The real problem here is that this is going to be a constantly written
16 bytes of nvram.
This will _very_ very quickly kill any type of flash that I'm aware
of...  as such it would have to be something
along the lines of a battery backed few KB of ram, with write out to
flash on failure.
Not sure what the lifetime in write-cycles of the standard cmos nvram
is, but even if (being battery backed)
it would work, that's a real hard place to find any space...?

Maciej

On Sun, Jul 27, 2008 at 14:29, Daniel Phillips <phillips <at> phunq.net> wrote:
> Hi Timothy,
>
> On Sunday 27 July 2008 14:10, timothy norman huber wrote:
>> Daniel,
>>
>> Several weeks ago, during a discussion over coffee, you mentioned a
>> rather intriguing optimization- by writing the start of the forward
>> log chain to NVRAM rather than disk, you avoid a disk seek and write
>> transfer for each forward log chain.  i believe there was another
>> motivation that had to do with avoiding traversing the entire volume
>> in case of a crash.  Was this design feature to overcome lame fsck
>> performance?
>
> The same motivation, actually.  Writing the start of a forward log
> chain to nvram instead of to some known location on disk means that I
> do not have to worry about doing unnatural things to optimize such
> loads as O_SYNC writing, where each transaction has to complete

(Continue reading)