[Leaplist] LINUX memory shortfall (long)
Bryan J. Smith
b.j.smith at ieee.org
Thu Jan 3 18:36:45 GMT 2008
Kevin Anderson wrote:
> Or am I OK with either AMD or Intel if it's a 64 bit OS?
> (It'll be ESX 3).
Define "OK"?
I have my own, professional opinions, but my point here wasn't to
debate AMD v. Intel. My point was to say there are differences
between AMD and Intel.
AMD x86-64 (AMD64) has units Intel IA-32e (EM64T) doesn't have.
That means there are additional considerations for both AMD64 and
EM64T you need to be aware of, in both regards. Assumptions are the
only problem.
E.g., Linux kernels that detect EM64T processors put the kernel into
a mode that uses "bounce buffers" below 4GiB below for all I/O. In
other words, you still have some 32-bit limitations.
At the same time, AMD64 processors can have "troublemaker" hardware
that have issues with its I/O MMU (EM64T doesn't have one). If you
disable portions (or the entire) I/O MMU, it falls back into
EM64T-like mode.
There are also cache consideratoins. AMD64 uses 128KiB L1 (4x
Intel), but smaller L2, and the new quads new an unified L3, although
AMD64 can load L1, L2 and L3 independently (crossbar interconnect).
EM64T uses 32KiB (1/4th AMD), very large L2, a few with L3 options
(chipset/mainboard), but anything in L1 must be in L2 (and anything
in L2 in L3 if it exists), because it's a simple bridge/bus design.
Latency v. throughput and other considerations for cache, I/O, etc...
come down to application, as well as compatibility, issues, etc...
Also note you will have library v. program compatibility issues with
distributions that ship with x86-64 kernels and libraries. Again,
x86-64 (48-bit "Long Mode" addressing) programs cannot use i386
(48-bit "segmented" addressing) libraries and vice-versa. If you've
never experienced it, please read my blog article on "Long Mode" for
some analogies and history on memory models.
You don't just "plop down" x86-64 unless you know what you're doing.
Although running an i386 Guest VM on a x86-64 ESX server is pretty
painless, and often recommended.
--
Bryan J. Smith Professional, Technical Annoyance
b.j.smith at ieee.org http://www.linkedin.com/in/bjsmith
------------------------------------------------------
Fission Power: An Inconvenient Solution
More information about the Leaplist
mailing list