Thanks to:
Dr. Thomas Blinn
Randall Cable
Werner Hahling
Blair Phillips
My original question is below. Indeed, if you fill four memory banks
with the same size, you will get four-way memory interleaving a
performance benifit. Tom pointed out that this was very much
application dependent and platform specific. I enclose his reply:
If you fill all four banks with the same sized memory, then when
the console firmware configures the memory, it will configure it
four-way interleaved. This will give better performance in some
memory access patterns. Whether the difference could be measured
with your work load on your system, I can't say. If you have got
well-behaved applications, then MOST of the "memory accesses" are
going to be from per-CPU cache anyway, although I/O transfers are
typically done directly to/from physical memory, and are so large
that they would (at least in principle) benefit from four-way vs.
two-way or non- interleaved memory. The EXACT performance figures
are platform specific, and I don't know if they're even published
anywhere you could easily find them.
Randall sent me a complete description of the configuration rules for
ES40 memory including details about mixing stacked and monolythic
technology DIMMs.
Werner pointed out that you can get either 2-way interleaving or 4-way
interleaving, i.e. three memory banks with the same memory will give
2-way interleaving and one separate. He wrote:
Interleaving can give quite some increase in memory access speed since
256bit reads can be performed in one go rather than in 4 reads if 1-way
interleaved or 2 reads if 2-way interleaved.
Blair pointed out that there was an issue about stacked vs unstacked
DIMMs.
> I want to add memory to my ES40 Model 1 AlphaServer which now has
> two banks of 1024MB each. There are two remaining banks. Does it
> make any difference in performance if I add another 2x1024MB or
> 1x2048MB? I would prefer the latter since this leaves me a spare
> bank for a future upgrade. But if the performance is improved by
> interleaving memory of the same size, I would do that.
>
> TIA,
>
> Peter Stern
> Chemical Physics Department
> Weizmann Institute of Science
> 76100 Rehovot, ISRAEL
>
> email: Peter.Stern_at_weizmann.ac.il
> phone: 972-8-9342096
> fax: 972-8-9344123
>
>
Received on Wed Aug 22 2001 - 05:59:13 NZST