| 1 |
Gavin Seddon, mused, then expounded: |
| 2 |
> |
| 3 |
> I have found Gentoo to be excellent due to it's configurability (is this |
| 4 |
> a word) and speed. I have no intention of changing now however, I am |
| 5 |
> always 'on-the-lookout for a superior workstation so I may change in the |
| 6 |
> future to ia64 Gentoo. Are Iitanium machines the best (robust/speed). |
| 7 |
> Or, are the Opterons or athlon64 better. For that matter how do amd |
| 8 |
> machines differ? |
| 9 |
> |
| 10 |
|
| 11 |
It depends. In general, looking at MTBF numbers, Itanium systems are |
| 12 |
less reliable than most 2P to 4P x86 servers out there. But more reliable |
| 13 |
than most RISC based servers, though IBM's Blue Gene blade servers may |
| 14 |
have an equivlant record or even be better. (Note: Realibility when I |
| 15 |
use it means 24x7 full time compute load, not shutting it down at night |
| 16 |
nor the system standing idle.) |
| 17 |
|
| 18 |
As to performance, it kind of depends. The bone stock Intel chipset |
| 19 |
Itaniums have fairly constrained FSB to memory bandwidth. Itanums where |
| 20 |
companies have used their own chipsets - such as us (SGI) and higher end |
| 21 |
HP, among others, can deliever some impressive bandwidth to memory and |
| 22 |
I/O. Thus improving overall computing throughput and high sustained |
| 23 |
compute power. |
| 24 |
|
| 25 |
And depending upon the problem set, an Itanium based system may be faster |
| 26 |
than amd64 based systems. If your problem set is more integer based or |
| 27 |
transaction based, the amd64 will perform better. If it's more FP based, |
| 28 |
memory constrained, threaded, and I/O bound, then an Itainium based system |
| 29 |
might be better. |
| 30 |
|
| 31 |
If you need more than 1 TB physical memory and up to 512 cpus in a single |
| 32 |
system image, then amd64 solutions won't be there until HyperTransport III, |
| 33 |
sometime next year, perhaps. |
| 34 |
|
| 35 |
Downside - you have to recompile your apps to get the best out of them. Any |
| 36 |
x86 32bit code will run in an emulation layer, which can be very fast or |
| 37 |
very slow. Setiathome, 32 bit client screamed on our Itanium's (ran it |
| 38 |
on a 512P system). Oh, and you'll need Intel's compiler. |
| 39 |
|
| 40 |
On the AMD side, figure about 80% FP performance and 102% Integer performance |
| 41 |
compared to an Itanium cpu. Memory bandwidth tends to match when AMDs are |
| 42 |
used in 4P or greater configs. I/O bandwidth is a bit constained in some |
| 43 |
instances, but seems pretty tolerable and really isn't much of an issue |
| 44 |
until more than 4 PCI-X slots are needed. But a lot of that is based |
| 45 |
on which chipset is being used. And to get the est out of your apps, you'll |
| 46 |
need to compile them with the PathScale compiler. |
| 47 |
|
| 48 |
If you're interested in an in-office/deskside style system. the AMD solution |
| 49 |
is probably a better match, especially if your application needs only 1 or |
| 50 |
2 Gfx pipes, and fits in less than 16 GB memory. |
| 51 |
|
| 52 |
If you need department scientifc serving, then benchmarking your apps on each |
| 53 |
platform is the better way to find which fits better. |
| 54 |
|
| 55 |
If you need serious visualizaton ability, with 3 or more - up to 16, Gfx |
| 56 |
pipes, then there isn't much choice, no AMD vendor has produced a system |
| 57 |
capable of doing that, though with some special Nvidia cards, it's possible |
| 58 |
to get 4 Gfx pipe on 2 PCIe x16 slots. That's not to say it works well, as it's |
| 59 |
pretty much beta hardware and drivers at the moment. |
| 60 |
|
| 61 |
Bob |
| 62 |
-- |
| 63 |
- |
| 64 |
QA Curmudgeon. |
| 65 |
Wacky and bizarre testing(TM) performed while-U-wait. |
| 66 |
- |
| 67 |
See http://www.gnu.org/philosophy/no-word-attachments.html |
| 68 |
- |
| 69 |
-- |
| 70 |
gentoo-amd64@g.o mailing list |
Archives