[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: ext3-2.4-0.9.5



On 31 Jul 2001 01:18:49 +1000, Andrew Morton wrote:
> The external journal code seems to work OK - brief usage details are at the
> web site.  The intent here is that the external journal be an NVRAM device
> (or a disk) which can be used to accelerate full-data journalling. 
> Simulation using a normal RAM drive indicates that we can double throughput
> with some loads (dbench) but not others (synctest).  More work is needed to
> fully characterise this.

Definitely.

> The simulator launches a (large) number of sub-processes.  Each
subprocess
> does the following:
> 
>   for 100 different filenames
>     create a file
>     write some data to the file (5k to 250k, exponential distribution)
>     optionally fsync() the file
>     close the file
>     optionally fsync() the file's parent dir
>     rename the file
>     optionally fsync() the file's parent dir
>     rename the file
>     optionally fsync() the file's parent dir
>     rename the file
>     optionally fsync() the file's parent dir
>     unlink the file from 30 passes ago.

Exim would have a roughly similar usage pattern to that, with fewer
renames :-)
The unlink might well be rather closer to the creation than your model
might put it - justification is that on many mail servers, particularly
on the multi-million-user set used by Freeserve et al in the UK, the
servers either take messages coming in from remote locations, put them
into a queue, then imediately deliver them to final resting place (for
Freeserve that would be on a big NAS user spool), or they accept local
messages and fire them off by SMTP.  In both cases the lifetime in the
MTA spool is somewhat under a second for 99% of traffic.  

This is why I am convinced that data journalling with the journal on a
NVRAM device, may well end up with fewer than 1% of transactions ever
actually hitting the disk (the majority will have lived and died before
the stuff gets flushed down to disk).

I'll think through your other comments later.... I'm just back from a
week away and am buried in both work and email.

	Nigel.







[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]