[rescue] Bus Speed v I/O rules of thumb?

[Francisco Javier Mesa-Martinez]

On Fri, 17 May 2002, Joshua D Boyd wrote:

> On Fri, May 17, 2002 at 12:23:54PM -0700, Francisco Javier Mesa-Martinez wrote:
> 
> > Of course a bus is a shared medium, so the more stuff you hang to it, the
> > more the performance is going to suffer. That is why putting 2 scsi boards
> > + 2 FDDI + a fast ethernet + .... on a single bus is not a good idea
> 
> I would expect that a SS10 level machine would be able to handle that sort
> of bus load.  25mhz*4bytes wide == 100megabytes/second== 800megabit/second. 
> Assuming that there is no DMA (is there DMA?  I don't remeber), then those
> devices will require 940megabits/second to run flat out.  Obvious problem,
> assuming no DMA and no bus overhear, we are still 140megabit short. In 
> reality, chances are unlikely that more than 2 network interfaces would be 
> running flat out, and so it is likely that at any given time, no more than 
> say 500megabits would be asked for from the machine, which shouldn't cause it 
> to break a sweat, unless the bus overhead is much worse than I would have
> expected (I expect about 25%).
> 
> At least, that is my thinking.  I could be wrong.[1]  

The problem is that you are thinking of a bus bandwith as a divisible
commodity, which would be somewhat true if the bus was fully switched. The
problem is as follows, most devices can not theoretically saturate the
bus, however every device does take the whole bus once it is writing to
it. Thus the behavior of the bus is no where to be as linear as your
calculations show... so besides bus overhead you also have to take into
account the actual inefficiency of the devices... several studies proved
that the behavior of a simple shared bus depended more on the slower
devices attached to buses than on the higher performance ones.

Also you have to take into account the nature of the traffic in the bus,
the requests follow a purely stochastic fashion thus contention is bound
to happen. And then you have to also model your devices as queues... and
then.... well moral of the story is that you can not model a bus by simply
aggregating the theoretical bw requirements for each device.

If you read any paper you will see that bus behavior is rarely linear, but
rather exponential (somewhat)....

 > -- 
> Joshua D. Boyd
> 
> [1] After several reedits to correct math errors, I realized that DMA isn't
>     usefull here, rather it would be something like I2O, which is unlikely
>     to be used.
> _______________________________________________
> rescue list - http://www.sunhelp.org/mailman/listinfo/rescue