1 |
"P.V.Anthony" <pvantony@×××××××××××.sg> posted |
2 |
46B0A3D5.6080203@×××××××××××.sg, excerpted below, on Wed, 01 Aug 2007 |
3 |
23:16:37 +0800: |
4 |
|
5 |
> Now the question is, is it better to go with 2 core or 4 core? |
6 |
> |
7 |
> The reason for this question is, that I heard there is a diminishing |
8 |
> return with more cores. Not sure if this is true with kernel 2.6.21 and |
9 |
> running at 64bit. |
10 |
> |
11 |
> The server needed to built is for the following apps. |
12 |
> |
13 |
> Hardware. |
14 |
> 1. Tyan Tank GT20 (B5191) |
15 |
> 2. 2 x Sata drives |
16 |
> 3. Software raid 1 |
17 |
> 4. 4Gb ram ecc |
18 |
> 5. Intel Core2Duo E6420 or Intel Core2Quad E6600 |
19 |
> |
20 |
> Apps. |
21 |
> 1. Gentoo linux 64bit |
22 |
> 2. Apache 2 |
23 |
> 3. MySql |
24 |
> 4. Postgres |
25 |
> 5. Qmail |
26 |
> 6. Pure-ftpd |
27 |
> 7. Mod_perl |
28 |
> 8. php |
29 |
> 9. ruby |
30 |
> |
31 |
> Will all the instances of the apps be shared among the cores? |
32 |
> |
33 |
> Please share the comments. |
34 |
> |
35 |
> I would really like to save some money. If the 2 core can do the job |
36 |
> there is quite a bit of savings buying just the 2 core. |
37 |
|
38 |
Your application is server, and others have dealt with it. I'll answer |
39 |
for desktop/workstation use, however, since that's what I'm doing here, |
40 |
and other readers may be interested. |
41 |
|
42 |
I've been running a dual socket Opteron for some time now, originally |
43 |
with Opteron 242s. I've upgraded memory to 8 gig (2x2 gig for dual |
44 |
channel off of each CPU/memory-controller, NUMA so each CPU tries to |
45 |
localize memory access to its own 4 gig), and run a 4 spindle kernel RAID |
46 |
(1, 6, 0). /tmp (with /var/tmp a symlink to it) is on tmpfs, so making |
47 |
use of some of that 8 gigs of RAM. RAID-1 /boot, RAID-6 main system, |
48 |
RAID-0 striped for speed swap (4x4GB=16GB swap), ccache, $PORTDIR, and |
49 |
/usr/src. Thus, there's little needing upgrade on the general system. |
50 |
|
51 |
This was my first dual CPU system (before dual-cores came out), and I |
52 |
really appreciate the flexibility it gives me. That said, now that I'm |
53 |
used to it, the dual CPU (single cores) is definitely a bottleneck at |
54 |
times, and I could really use a quad system. That in fact is what I'm |
55 |
planning. I've money set aside for upgrading to dual dual-cores, maxing |
56 |
out the system with Opteron 290s, but at a going price of $1400+ for the |
57 |
pair on pricewatch.com, I'm hoping to wait until Barton comes out and |
58 |
that it will lower the price on the old socket 940 Opterons. I'm hoping |
59 |
it'll drop the pricepoint to about the current one of the 285s, saving me |
60 |
$300-400. |
61 |
|
62 |
I'm definitely looking forward to the upgrade, tho. It's nice to be able |
63 |
to run two CPU hogs, CPU affinity set so each is scheduling on its own |
64 |
CPU, without seriously affecting responsiveness. However, the two CPUs |
65 |
is definitely a limiting factor, now, particularly since other than |
66 |
memory, the rest of the system's all attached to socket 0, so apps such |
67 |
as X run more efficiently on it. By upgrading to dual dual-cores, I'll |
68 |
have two cores able to schedule on that direct-connected socket 0, plus |
69 |
two cores on the less well connected socket 1. Where now I can run X on |
70 |
CPU 0 and my other CPU hogging app on CPU 1, I'm quite looking forward to |
71 |
being able to run both on separate cores on socket 0, while the two cores |
72 |
on socket one could be doing a full emerge -e world for example, without |
73 |
affecting my CPU intensive interactive stuff on the two socket 0 cores |
74 |
/at/ /all/. |
75 |
|
76 |
So yes, I appreciate the dual CPU, but while the popular press is still |
77 |
asking if desktop tasks can scale to dual-core, let alone quad-core, I'm |
78 |
on dual-CPU right now and definitely could use the quad-core or 2 by dual- |
79 |
core system right now! I can't at this point say that I'd be able to |
80 |
make good use of 8 cores, but four, certainly, as two is feeling a bit |
81 |
constrained at this point. |
82 |
|
83 |
As for Intel vs. AMD, with the hypertransport interconnect system AMD |
84 |
has, with AMD's hardware IOMMU for DMA access above the 4 gig boundary |
85 |
(Intel has to emulate it with software bounce-buffers on most of its |
86 |
hardware, kinda nullifying the point of DMA), and with the vaunted Intel |
87 |
performance lead mostly a 32-bit phenomenon, dual dual-core AMD is often |
88 |
more efficient on 64-bit than quad-core Intel. |
89 |
|
90 |
The biggest down side for AMD here is the severe lack of open source |
91 |
support for anything faintly modern, GPU-wise. Unfortunately, Intel's |
92 |
the only real option for anything even half modern in terms of open |
93 |
source driven graphics chipsets. For an x86_64 Linux desktop/ |
94 |
workstation, that more than cancels out the lead AMD would arguably have |
95 |
in multi-core/multi-CPU support, and if I were buying now, it'd be Intel |
96 |
for that reason. AMD simply has to get its act together, and release at |
97 |
least enough specs for its ATI graphics product so the community can |
98 |
build the drivers itself. Or sponsor its own open source drivers. |
99 |
Either way. |
100 |
|
101 |
-- |
102 |
Duncan - List replies preferred. No HTML msgs. |
103 |
"Every nonfree program has a lord, a master -- |
104 |
and if you use the program, he is your master." Richard Stallman |
105 |
|
106 |
-- |
107 |
gentoo-amd64@g.o mailing list |