On Sunday 13 March 2011, Dirk Behme wrote:
> Hi Arnd,
>
> reading the flashbench README [1] what I really like is the
> explanation of the -a and the -O option. Explaining the options in
> this README, giving the examples and how to interpret the resulting
> figures really does help.
>
> Somehow I miss something similar for the -s and -f options. I.e. how
> to select the proper values for --scatter-order and --scatter-span and
> how to interpret the output of -s and -f.
>
> Once I understood it, I would be able to send a patch for the README ;)
>
> Additionally, it would be nice to give the flashbench options used for
> the graphs [2] [3] in the LWN article. The LWN article explains quite
> nicely how to interpret the given graphs, but it's not mentioned which
> flashbench options were used to get these graphs.
>
> Many thanks for your help and best regards
>
> Dirk
>
> [1]
> http://git.linaro.org/gitweb?p=people/arnd/flashbench.git;a=blob;f=README;h…
>
Sorry for the late reply. I promise I'll get to it and update the
README.
I should actually remove --scatter-order, it's too difficult to
understand this. It specifies the log2 of the number of blocks
in terms of --blocksize to be tested at the start of the medium.
The output file can be interpreted by
gnuplot -p -e 'plot "output.file"'
or by importing it into a spreadsheet program like oocalc and
using the XY chart function on two columns.
For --findfat, the output shows how each of the first N erase blocks
on the drive reacts to certain access patterns within the erase
block. Most drives do something different for a few blocks in the
beginning to optimize storing the FAT on them. Each column is one
erase block here. If they are all the same, the card does not have
an optimzied FAT area.
Arnd
On Friday 04 March 2011 20:16:05 Xianghua Xiao wrote:
> From the data below it appears APACER has a 1MB erasing block instead
> of the typical 4MB?
Yes, and this is not surprising for an SLC card, although it is
the first one I have seen.
> Also from the following email(waiting for approval on the list, will
> forward to you soon), that looks like a 4MB erasing block but is
> slower than APACER.
Sorry, I need to find the place on the mailing list settings to allow
non-members to post, and approve the mails you already sent.
> Does open-au-nr with higher number mean the underlying filesystem esp
> ext4 will do better,
Yes
> does that imply multithread parallel writing to some extent?
No, it's not about threads, but about how the data is laid out.
ext4 will have to write data, metadata and journal data for
many accesses, but these three are normally in different locations
on the driver, so you need at least three open segments for the first
process that is writing data. If you have other processes that also
write to the drive, or one process writing to multiple files, you
will need more.
> open-au-nr means how many different segments you can use
> to write to in the same time(so you don't need wait for one segment)?
Yes. The apacer card evidently can write to three segments, but not
to six. Can you also try 4 and 5 segments? The interesting number
is the maximum.
For the unigen card, it looks like it does not handle random access
well, independent of the number of AUs. Best try again without --random.
> I probably will recommend APACER over UNIGEN, does that make sense?
Depends on the other measurements.
Arnd