1 |
P.V.Anthony, mused, then expounded: |
2 |
> |
3 |
> I must say that I love the intel drivers for the network and sata. If |
4 |
> only amd has some chip sets for their own cpus and good open source |
5 |
> drivers for their chip sets. Intel has that covered and I think the |
6 |
> intel drivers are open source. |
7 |
> |
8 |
|
9 |
It's more the chipset than the drivers. The Intel GigE chip controller |
10 |
has a built-in tcp offload engine that removes cpu loading of network |
11 |
traffic, if enabled in the Linux kernel. |
12 |
|
13 |
Of the available chipsets for AMD support, only Nvidia has a proprietary |
14 |
driver, thus the need for the reverse engineered driver. Most non-Nvidia |
15 |
chipset platform for AMD use either Broadcom or Intel GigE chips. And the |
16 |
Broadcom drivers have been well tested over the years. So, other than |
17 |
TCP offload, I wouldn't avoid evaluating a specific option due to a concern |
18 |
about drivers. |
19 |
|
20 |
SATA - again with AMD you have some options - nVidia or Broadcom, either work fine, |
21 |
as does the Intel SATA controller. That said, be cautious of using the Intel |
22 |
software raid from the bios. I've not found it to be consistantly implemented, |
23 |
nor for it to always work well. Generally, unless a real hardware raid is offered, |
24 |
such as a LSI 1064 or 1068 SAS/SATA chip, stay with kernel supported software raid. |
25 |
|
26 |
And SATA in general, in rack mount enclosures, can suffer in performace, especially |
27 |
in 1U enclosures due to fan vibration. Inspect the fan mounting and insure that |
28 |
the fans are isolated from the chassis with rubber mounting bushings or grommits. |
29 |
|
30 |
And make sure the latest bios, bmc, and other firmware is installed and up to date! |
31 |
It's important. |
32 |
|
33 |
Don't confuse yourself by thinking that because Intel makes it's own chipset, that |
34 |
it's a better platform than anything available for AMD. Intel has two goals for |
35 |
making it's own chipset - vertical control of it's market, thus maximizing it's profit, |
36 |
and because OEMs don't want to have to think about having to do any real engingeering |
37 |
when putting their logo on a platform. AMD's being forced into the platform direction |
38 |
because of OEMs. |
39 |
|
40 |
ALso, consider that while Intel does have very good engineering, it's entire platform |
41 |
is proprietary. Only recently did it open up it's socket to counter AMD with |
42 |
programmable, pluggable ASICs. |
43 |
|
44 |
AMD on the other hand runs on an open bus architecture. So anyone can add things to it, |
45 |
with minimal licensing - making something for HyperTransport requires joining the |
46 |
consortium (and it's not part of AMD). So it open source is important, I'd think |
47 |
you'd want an open platform, where you get to choose the best infrastructure |
48 |
for your computing needs. |
49 |
|
50 |
Tyan offers both platforms to best meet what you think you're needs are. And you should |
51 |
look at those needs in that perspective. From your mix of apps, I'm not so sure the |
52 |
Intel platform is going to be the best. It really depends upon the overall mix. If |
53 |
you have more active database access than network traffic, then something with a higher |
54 |
sustained memory bandwidth, along with higher i/o bandwidth is going to provide much |
55 |
better performance. If the mix is mostly crunching on data retrieved, thus lots |
56 |
of FP or integer activity, than cpus with large caches and speed are more important. |
57 |
|
58 |
Typically, when compute power is more important to the mix, the Intel Core2 platform |
59 |
will perform the best. When memory and i/o bandwidth are important - database |
60 |
transactions, the AMD platform delivers a higher sustained bandwidth. And if |
61 |
the job mix is very peaky - lots of variations with long periods of lull time, |
62 |
the AMD platform will deliver overall power savings and require less cooling for |
63 |
a decrese in peak performance. |
64 |
|
65 |
Bob |
66 |
- |
67 |
-- |
68 |
gentoo-amd64@g.o mailing list |