Re: [PATCH] mmc: Added quirks for Ricoh 1180:e823 lower base clock frequency

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 12 July 2011 19:30:42 Chris Ball wrote:
> On Tue, Jul 12 2011, Manoj Iyer wrote:
> > btw only the 1st write was slower, subsequent writes looks ok.
> > [..]
> > I have attached the output of flashbench and the time test to
> >
> > http://launchpad.net/bugs/773524
> > [..]
> > == Finding the number of open erase blocks ==
> > u@u:~/flash/flashbench$ sudo ./flashbench -O --erasesize=$[4 * 1024 *
> > 1024] --blocksize=$[256 * 1024] /dev/mmcblk0  --open-au-nr=2
> > 4MiB    6.36M/s
> > 2MiB    6.24M/s
> > 1MiB    6.17M/s
> > 512KiB  6.19M/s
> > 256KiB  6.22M/s
> > u@u:~/flash/flashbench$
> > [..]
> > ====== AFTER PATCH ========
> > [..]
> > == Finding the number of open erase blocks ==
> > u@u:~/flash/flashbench$ sudo ./flashbench -O --erasesize=$[4 * 1024 *
> > 1024]         --blocksize=$[256 * 1024] /dev/mmcblk0  --open-au-nr=2
> > [sudo] password for u:
> > 4MiB    5.49M/s
> > 2MiB    6.22M/s
> > 1MiB    6.22M/s
> > 512KiB  6.21M/s
> > 256KiB  6.21M/s
> > u@u:~/flash/flashbench$
> 
> That's interesting.  Arnd, any idea why only the first test of the
> flashbench run would be slower after the patch?

I would very much expect that to be nonreproducible. The first row
in each test is the result of a single write() system call and does
not get averaged out. More importantly the time for each write
depends a lot of the state of the card before the write.

For instance when you do a lot of random writes to a card, optionally
take it out and put it into a different machine, and then do a large
linear write, that linear write will be very slow because the
card has to garbage collect all the random writes that were done
earlier. After a few writes (usually one is enough), it gets back
to the full performance.

	Arnd