Strange NOR read/write throughput behaviour

Strange NOR read/write throughput behaviour

[Bernhard Priewasser]

Hi all,

can anybody help me explaining this diagram of NOR read throughput?

	http://files.lamark.de/nor_read.png

As soon as a file, read from a StrataFlash containing only this file,
exceeds erase block size (128KiB), file reading lasts a portion of time
longer and read throughput stagnates...

Write throughput seems to be constant around erase block size:

	http://files.lamark.de/nor_write.png

Best regards,
Bernhard

Re: Strange NOR read/write throughput behaviour

[Jared Hulbert]

On 11/15/05, Bernhard Priewasser <priewasser@gmail.com> wrote:
> Hi all,
>
> can anybody help me explaining this diagram of NOR read throughput?
>
>         http://files.lamark.de/nor_read.png
>
> As soon as a file, read from a StrataFlash containing only this file,
> exceeds erase block size (128KiB), file reading lasts a portion of time
> longer and read throughput stagnates...
>
> Write throughput seems to be constant around erase block size:
>
>         http://files.lamark.de/nor_write.png
>
> Best regards,
> Bernhard
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>


Could you please explain the experiment a little more?

How are you measuring read and write throughput?

Which StrataFlash are you using?  What board are you connecting the
flash to?  What kernel and MTD are you using?

Re: Strange NOR read/write throughput behaviour

[Bernhard Priewasser]

 > Could you please explain the experiment a little more?
Sure!

 > How are you measuring read and write throughput?
C-Program, gettimeofday(). Very clean conditions: Empty flash, mount 
JFFS2, fopen one file. Measure duration of fwrite(filesize). Throughput 
= 1/duration.

 > Which StrataFlash are you using?
E28F128J3, 128 Mbit. 128KiB erase block size.

 > What board are you connecting the flash to?
Hynix HMS30C7202 evaluation board.

 > What kernel and MTD are you using?
Kernel 2.6.11-hnx3
MTD/JFFS2 of kernel 2.6.11. (I know, I should try newest CVS...)

-- 
Bernhard

Re: Strange NOR read/write throughput behaviour

[Jared Hulbert]

On 11/18/05, Bernhard Priewasser <priewasser@gmail.com> wrote:
>  > Could you please explain the experiment a little more?
> Sure!
>
>  > How are you measuring read and write throughput?
> C-Program, gettimeofday(). Very clean conditions: Empty flash, mount
> JFFS2, fopen one file. Measure duration of fwrite(filesize). Throughput
> = 1/duration.
>
>  > Which StrataFlash are you using?
> E28F128J3, 128 Mbit. 128KiB erase block size.
>
>  > What board are you connecting the flash to?
> Hynix HMS30C7202 evaluation board.
>
>  > What kernel and MTD are you using?
> Kernel 2.6.11-hnx3
> MTD/JFFS2 of kernel 2.6.11. (I know, I should try newest CVS...)
>
> --
> Bernhard
>


What about the read test.

Re: Strange NOR read/write throughput behaviour

[Bernhard Priewasser]

> What about the read test.

As simple as you can imagine. Write a file (fopen, fwrite, fclose), wait 
some time, read the file (fopen, fread, fclose).
Do you think there are caching mechanism responsible for it?

-- 
Bernhard

Re: Strange NOR read/write throughput behaviour

[Jared Hulbert]

On 11/22/05, Bernhard Priewasser <priewasser@gmail.com> wrote:
> > What about the read test.
>
> As simple as you can imagine. Write a file (fopen, fwrite, fclose), wait
> some time, read the file (fopen, fread, fclose).
> Do you think there are caching mechanism responsible for it?
>
> --
> Bernhard
>

Sorry for the long wait.

Yes, It sure looks like a caching effect.  What is happening at
128KiBs?  Let me ask around a little more.  This is an arm7 processor,
no L1 or L2 caches?

Re: Strange NOR read/write throughput behaviour

[Bernhard Priewasser]

> Sorry for the long wait.
No problem.

> Let me ask around a little more.  This is an arm7 processor,
> no L1 or L2 caches?
ARM 720T, 8 KiB combined instruction/data cache. I don't know if there 
is any L2-Cache on Hynix HMS30C7202; but if there is one, I doubt it's 
as big as 128 KiB.
It looks pretty much like there is a cache sized 128 KiB somewhere 
beetween fread() and the flash. But don't know where...

-- 
Bernhard