Re: Memory on PCI

["Brian A. Cole" <cole@nevis1.nevis.columbia.edu>]

On Thu, 13 Feb 1997, John R Pierce wrote:


> >   Thanks for the info. I hadn't thought about trying to adapt a video
> > card, but maybe it's not such a bad idea.
> 
> I suppose I should ask why you aren't just using main memory?  Its at least 2X
> as fast on typical triton based pentium systems (PCI bus has maybe 80-100MB/sec
> useable burst bandwidth, EDO main memory on tritons has 200MB/sec or better
> bandwidth, so even hammering it at burst PCI speeds only cuts into the
> CPU-memory bandwidth by a percentage)
> 
> -jrp
> 

HI John,

  Well, this is a good question. We have an unusual application in 
which we need to shove relatively large blocks of data (4K-64K)
into memory, manipulate a small fraction of the data, and then 
send it back out over ATM. Since our input device and the ATM interface 
have intelligence and could act as masters on PCI, most of the
handling of the data could take place w/o crossing the host
bridge. Since our rates will automatically be limited by the PCI
bus anyway, it seemed reasonable to dump the data into PCI memory
if it were readily available. This would let the operating system 
and other parts of our apllication code work with the main memeory
unimpeded. Also, since our application will not make very effective use 
of the cache, clobbering the cache with data that would most likely 
never get re-fetched did not seem useful. 

Clearly we can get by using main memory because, as you point out,
it has sufficiently high bandwidth, but we have to now pay more 
attention to whether or not the operating system can hog access to
memory for periods longer than our maximum latency. 

It's probably clear to you that we are not PCI experts. Our background
is more in VME systems, so we're still climbing the learning curve
on how to do what we want in PCI

Cheers,

Brian

���