[Leaplist] Intel Yonah Mini-ITX and new Atom (ultra-low'n slow x86) -- WAS: RFC: LEAP Linux NON-dream boxes

[John Simpson]

On 2008-03-03, at 1603, Bryan J. Smith wrote:
>
> More interestingly is that Intel will start embedded Atom into its
> chipsets.  I.e., TCP/IP and Storage off-load.  So we may soon see
> low-end, server-class GbE and RAID-5/6 off-load.  ;)

just as an interesting side-note, to explain the term "off-load" to  
people who may not be familiar with it...

when i started working for KUA (back in 2000) the guy who ran their AS/ 
400 told me that it ran on a "power pc" chip, which was the same thing  
they were calling a "G5" in the mac world.

when i asked him what the secret was, how come an AS/400 supposedly  
ran so much faster than a mac, he explained that the architecture was  
totally different. there was one primary CPU (which could be expanded  
to two, four, eight, or more CPUs if needed) for doing the actual job  
of the machine, but then there was a separate CPU for managing the I/O  
bus, and a separate CPU for managing each ethernet (and token ring)  
card, and a separate CPU to deal with managing the disks, and so  
forth... he said that all told, there were about 15 CPUs inside the  
case, whose only jobs were to offload the background processing from  
the main CPU, and that each one of these "child" CPUs was also a power  
pc chip of some kind (maybe not G5 class, but still an entire CPU with  
its own miniature "firmware" to run, and its own RAM, and so forth.)

so basically the difference between that kind of "mainframe" system,  
and a desktop system, was that on the larger boxes, IBM off-loaded the  
normal "background tasks" to multiple other CPUs, leaving the main CPU  
free to concentrate on nothing but the primary job of the machine.

that's what "off-loading" is... moving the work away from the primary  
CPU, so that it's free to take care of YOUR program.

of course the OS and drivers will need to be re-written in order to  
take advantage of this, but i think the drivers will become simpler-  
the card will probably have an amount of RAM (16MB maybe?) which can  
be addressed by the CPU, the card will just write any incoming packets  
to its own RAM, and the card will then tell the kernel where in RAM to  
find those packets. the whole network I/O becomes a bunch of memory- 
copy operations (which are fast, especially on a fast bus) rather than  
dedicated I/O operations (which aren't nearly as fast.)

the same applies to hard drives- if the kernel is able to "teach" the  
hard drive controller how to handle a lot of the low-level details,  
maybe even as high as handling the filesystem operations, then a  
syscall like open() could be handed directly to the disk controller,  
and IT would do the work, leaving the main CPU to handle other  
processes. it may not make much of a difference for a desktop machine,  
but for a server which is handling a lot of users' processes, or for a  
machine which is doing hardcore number-crunching, it can be a very  
noticeable effect- like seemingly doubling the speed of the machine.

--------------------------------------------------------
| John M. Simpson  --  KG4ZOW  --  Programmer At Large |
| http://www.jms1.net/                 <jms1 at jms1.net> |
--------------------------------------------------------
|   Hope for America  --  http://www.ronpaul2008.com/  |
--------------------------------------------------------





-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 194 bytes
Desc: This is a digitally signed message part
Url : http://lists.leap-cf.org/pipermail/leaplist/attachments/20080304/80bfe5ce/PGP.bin