Re: file system for solid state disks

["Daniel J Blueman" <daniel.blueman@gmail.com>]

On 23/08/07, Fengguang Wu <wfg@mail.ustc.edu.cn> wrote:
> On Thu, Aug 23, 2007 at 01:56:17PM +0100, Daniel J Blueman wrote:
> > Hi Fengguang,
> >
> > On 23/08/07, James Courtier-Dutton <James@superbug.co.uk> wrote:
> > > Daniel J Blueman wrote:
> > > > On 23 Aug, 07:00, Jan Engelhardt <jengelh@computergmbh.de> wrote:
> > > >> On Aug 23 2007 01:01, Richard Ballantyne wrote:
> > > >>
> > > >>> What file system that is already in the linux kernel do people recommend
> > > >>> I use for my laptop that now contains a solid state disk?
> > > >>>
> > > >> If I had to choose, the list of options seems to be:
> > > >>
> > > >> - logfs
> > > >>   [unmerged]
> > > >>
> > > >> - UBI layer with any fs you like
> > > >>   [just a guess]
> > > >>
> > > >> - UDF in Spared Flavor (mkudffs --media-type=cdrw --utf8)
> > > >>   [does not support ACLs/quotas]
> > > >
> > > > Isn't it that with modern rotational wear-levelling, re-writing hot
> > > > blocks many times is not an issue, as they are internally moved around
> > > > anyway? So, using a journalled filesystem such as ext3 is still good
> > > > (robustness and maturity in mind). Due to lack of write buffering,
> > > > perhaps a wandering log (journal) filesystem would be more suitable
> > > > though? I use ext3 on my >35MB/s compact flash filesystem.
> > > >
> > > > I can see there being advantage in selecting a filesystem which is
> > > > lower complexity due to no additional spatial optimisation complexity,
> > > > but those advantages do buy other efficiency (eg the Orlov allocator
> > > > reducing fragmentation, thus less overhead), right?
> > > >
> > > > Also, it would be natural to employ 'elevator=none', but perhaps there
> > > > is a small advantage in holding a group of flash blocks 'ready' (like
> > > > SDRAM pages being selected on-chip for lower bus access latency) -
> > > > however this no longer holds when logical->physical remapping is
> > > > performed, so perhaps it's better without an elevator.
> > > >
> > > > Clearly, benchmarks speak...but perhaps it would make sense to have
> > > > libata disable the elevator for the (compact) flash block device?
> > > >
> > > > Daniel
> > >
> > > Also, sector read ahead will actually have a performance impact on
> > > Flash, instead of speed things up with a spinning disc.
> > > For example, a request might read 128 sectors instead of the one
> > > requested at little or no extra performance impact for a spinning disc.
> > > For flash, reading 128 sectors instead of the one requested will have a
> > > noticeable performance impact.
> > > Spinning discs have high seek latency, low serial sector read latency
> > > and equal latency for read/write
> > > Flash has low seek latency, high serial sector read latency and longer
> > > write than read times.
>
> A little bit of readahead will be helpful for flash memory.  Its latency is
> low, but sure not zero. Asynchronous readahead will help to hide the latency.
>
> > I was having problem invoking the readahead logic on my compact flash
> > rootfs (ext3) with tweaking the RA with 'hdparm -a' from 8 to 1024
> > blocks and some benchmarks (I forget which).
> >
> > Fengguang, what is your favourite benchmark for finding differences in
> > readahead values (running on eg ext3 on a flashdisk), with the current
> > RA semantics in mainline kernels (eg 2.6.23-rc3)?
>
> My favorite test cases are
>
> big file:
>         time cp $file /dev/null &>/dev/null
>         time dd if=$file of=/dev/null \ bs=${bs:-4k} &>/dev/null
>
> big file, parallel:
>         time diff $file $file.clone
>
> small files:
>         time grep -qr 'doruimi' $dir 2>/dev/null
>
> Don't forget to clear the cache before each run:
>         echo 2 > /proc/sys/vm/drop_caches

The maximal case we're looking for is where up to 1024-block
read-ahead doesn't pay off, but actually wastes finite bandwidth, thus
time.

We clearly need a database-type workload which forward reads enough to
open the RA window some, but then reads at a different location. We
then prove that we benefit with a smaller read ahead window, at
negligible cost to your linear case.

To open the RA window, I know we need no competing threads, but how
far do we need to sequentially read? I'll cook a micro-benchmark with
memory-backed files.
-- 
Daniel J Blueman