I/O scheduling (pdflush) on the XO

[Mitch Bradley]

JFFS2 does automatic compression so every write invokes zlib 
compression.  That is a rate-limiting factor on this machine - the raw 
NAND write speed is several times faster than the compress speed.

For files with large pieces, that compression results in a 2x space 
savings.  But JFFS2 can behave quite badly when presented with lots of 
small writes.  In the limiting case of one-byte writes, JFFS2 can store 
65 bytes for every byte presented (the base node size is 64 bytes).

M. Edward (Ed) Borasky wrote:
> Mikus Grinbergs wrote:
>   
>> For the fun of it, started a resource-intensive task (100% CPU, 
>> sporadic floods of disk writes) on my G1G1 XO.  Eventually, it 
>> failed with a "severe (38)" fortran error trying to write its 
>> checkpoint file {mikus note: many quick small write operations}.
>>     
>
> Why? This device has a 433 MHz 32-bit processor, 256 MB of RAM and a 1 
> GB jffs2 flash disk. What do you expect?
>   
>> That situation may be a pathological one -- but what I found 
>> interesting was that in again doing something similar (different 
>> worktask), 'top' showed 'pdflush' requiring lots of CPU (up to 
>> 52%!!). [And 'jffs2_gcd_mtd0' keeps showing up on 'top', too.]
>>
>> I take it that my XO is *struggling* when presented with heavy 
>> "disk I/O".
>>     
>
> Again, where's the surprise here? An XO is *not* a scientific 
> workstation! It's a computer for elementary school students!
>
> If you want to run stuff like this, do it on a "school server" class 
> machine and write an XO client.
> _______________________________________________
> Devel mailing list
> Devel at lists.laptop.org
> http://lists.laptop.org/listinfo/devel
>