Re: Solaris ZFS on Linux [Was: Re: the " 'official' point of view" expressed by kernelnewbies.org regarding reiser4 inclusion]

[Nate Diller]

On 7/31/06, David Lang <dlang@xxxxxxxxxxxxxxxxxx> wrote:
> On Mon, 31 Jul 2006, Nate Diller wrote:
>
> >
> > On 7/31/06, Matthias Andree <matthias.andree@xxxxxx> wrote:
> >> Adrian Ulrich wrote:
> >>
> >> > See also: http://spam.workaround.ch/dull/postmark.txt
> >> >
> >> > A quick'n'dirty ZFS-vs-UFS-vs-Reiser3-vs-Reiser4-vs-Ext3 'benchmark'
> >>
> >> Whatever Postmark does, this looks pretty besides the point.
> >
> > why's that?  postmark is one of the standard benchmarks...
> >
> >> Are these actual transactions with the "D"urability guarantee?
> >> 3000/s doesn't look too much like you're doing synchronous I/O (else
> >> figures around 70/s perhaps 100/s would be more adequate), and cache
> >> exercise is rather irrelevant for databases that manage real (=valuable)
> >> data...
> >
> > Data:
> >       204.62 megabytes read (8.53 megabytes per second)
> >       271.49 megabytes written (11.31 megabytes per second)
> >
> > looks pretty I/O bound to me, 11.31 MB/s isn't exactly your latest DDR
> > RAM bandwidth.  as far as the synchronous I/O question, Reiser4 in
> > this case acts more like a log-based FS.  That allows it to "overlap"
> > synchronous operations that are being submitted by multiple threads.
>
> what you are missing is that apps that need to do lots of syncing (databases,
> mail servers) need to wait for the data to hit non-volitile media before the
> write is complete. this limits such apps to ~1 write per revolution of the
> platters (yes it's possible for a limited time to have multiple writes to
> different things happen to be on the same track, but the counter is the extra
> seek time needed between tracks)

this is true so long as there is only one thread submitting I/O and
doing fsync().  for something like a mail server, it can run
multi-threaded, and still get data integrity, if the changes are
spread out across more than one file.

> so any benchmark that shows more transactions then the media has revolutions is
> highly suspect (now if you have battery-backed cache, or the equivalent you can
> blow past these limits)

not all workloads are completely serial, transactions themselves may
have no inter-dependencies at all.  so it depends on the benchmark,
and what workload you're measuring.  in cases like this, threading can
have a big advantage.

NATE
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/