I CAN see how this is happening.
I am not sure of the design of the UP2000, or XP1000, but I know that
other alpha platforms use something called "memory interleaving".
Memory references are spread out across DIMMs. This allows more
fetches to be done in parallel. Since there is often a high initial
cost for grabbing a piece of data off of the DIMM, this can be a big
win.
Once that first piece of data is grabbed, all the other data is ready
and waiting.
For example, let's presume it costs 80 cycles to get a piece of data
ready for consumption, and let's say that I have 8 pieces of data that
I want to retrieve.
If I can do 4 in parallel: (with four DIMMS)
I wait:
80 cycles for the first piece.
0 cycles for the second peice.
0 cycles for the third piece.
0 cycles for the fourth piece.
80 cycles for the fifth piece.
0 cycles for the sixth piece.
0 cycles for the seventh piece.
0 cycles for the eigth piece.
If I can do 8 in parallel: (with eight DIMMS)
I wait:
80 cycles for the first piece.
0 cycles for the second peice.
0 cycles for the third piece.
0 cycles for the fourth piece.
0 cycles for the fifth piece.
0 cycles for the sixth piece.
0 cycles for the seventh piece.
0 cycles for the eigth piece.
With 8 DIMMS, I've rid myself of an extra 80 cycles.
(These numbers are by no means exact, but the picture they paint is correct.)
This win will really show up with programs that stream through
memory. (McCalpin Streams...)
General performance rule:
If you want maximum memory performance, fully populate your memory
banks with equally sized DIMMS.
--Phil
Compaq: High Performance Server Division/Benchmark Performance Engineering
---------------- Alpha, The Fastest Processor on Earth --------------------
Phillip.Ezolt@compaq.com |C|O|M|P|A|Q| ezolt@perf.zko.dec.com
------------------- See the results at www.spec.org -----------------------
On Fri, 23 Jun 2000, Rich Payne wrote:
> On Fri, 23 Jun 2000, W Bauske wrote:
>
> > Iwao MAKINO wrote:
> > >
> > >
> > > If you use dual CPU, you MUST fill 8 memory slot to get FULL performance.
> > > 667/4 is pretty good... 750/8 is even better
> > >
> >
> > Do the math. If you have 4 SDRAM DIMM's, each 8 bytes wide,
> > that's a 32 byte wide data stream, also called 256bits. And,
> > clocked at 83.33MHZ, you get 2.6GB/sec, on a single bank. That
> > is what the board is spec'd at. The UP2K uses the same memory
> > architecture as an XP1000, near as I can tell. One can order
> > memory for an XP1000 and it runs perfectly fine in a UP2K.
>
> UP2K takes PC100 registered DIMMS, I think XP1000 is the same.
>
> > Going to 8 DIMM's won't get you any more performance.
>
> Yeah, I can't see how this would make a difference either, unless you are
> increasing the memory size.
>
> --rdp
>
> > (I run both UP2K's and XP1000's)
> >
> >
> > Wes
> >
> >
>
> --
> Rich Payne
> rpayne@alphalinux.org www.alphalinux.org
>
> --
> To unsubscribe: send e-mail to axp-list-request@redhat.com with
> 'unsubscribe' as the subject. Do not send it to axp-list@redhat.com
>
>
>
-- To unsubscribe: send e-mail to axp-list-request@redhat.com with 'unsubscribe' as the subject. Do not send it to axp-list@redhat.com
This archive was generated by hypermail version 2a22 on Sat Jul 1 05:31:31 2000 PDT
Send any problems or questions about this archive to webmaster@alphalinux.org.