1 |
On Thu, Dec 02, 2004 at 12:21:29AM +0800, Zhang Weiwu wrote: |
2 |
|
3 |
> Okay, so I am wrong: is it true that U5 and U10 are worse than U2 when |
4 |
> equiped with 300~400 CPU? |
5 |
|
6 |
Usual disclaimers apply when discussing performance: YMMV, performance |
7 |
is highly workload-dependent, it's possible to contruct pathological |
8 |
workloads that prove almost any ridiculous contention, etc, etc. |
9 |
|
10 |
The best CPUs available for U5/10 (440) will probably offer you |
11 |
performance similar to some of the slower available for the U2 (200, |
12 |
250). |
13 |
|
14 |
The U5/10 are PC systems and they are full of bottlenecks. The memory |
15 |
subsystem is a joke. Most of the CPUs, even the higher-clocked ones, |
16 |
have inadequate cache. The U5 has no UPA slot so graphics will always |
17 |
be slow. The PCI in these systems is slower than a modern PC's and |
18 |
much slower than the U2's 64-bit SBUS. The standard disk system in |
19 |
the U5/10 is IDE - to make matters worse the chipset is a junk ALi one |
20 |
with severe bugs that sometimes make DMA unusable while the U2 has a |
21 |
fast/wide SCSI system which is certainly not very modern but |
22 |
infinitely better. |
23 |
|
24 |
A U5/10 is comparable in both performance and quality with an |
25 |
eMachines celeron system of similar clock and memory capacity, |
26 |
provided you can use IDE DMA on both. If the U5/10 cannot use DMA, it |
27 |
will be slower still. |
28 |
|
29 |
A U2 is comparable with a higher-clocked P3 system of similar memory |
30 |
capacity and number of CPUs. There are addons available for either |
31 |
type of system that could give it an advantage (for example, an U2 |
32 |
with FC disks would gain further substantial advantages over an |
33 |
IDE-equipped PC, but an Ultra2 SCSI-equipped PC would have an |
34 |
advantage over an U2 using the standard onboard SCSI subsystem). All |
35 |
U2s are SMP-capable, while most P3 systems are not. Tasks which are |
36 |
memory-bandwidth intensive will fare even better on the U2, which has |
37 |
about 4 times the bandwidth of common P2/P3 PC boards (the U2 memory |
38 |
system is 576 bits wide and runs at 100 MHz - similar in performance |
39 |
to modern Athlon and Opteron systems). |
40 |
|
41 |
> Another question I had for a long time: does SMP really makes sense on |
42 |
> desktop, browsing and word processing? I don't expect dual CPU average |
43 |
> performance 200% of single CPU, I don't even expect 120% performance, I |
44 |
> thought not needed for multi-process for desktops? |
45 |
|
46 |
If you never run an application that uses threads, or you never run |
47 |
more than one application at once, SMP will give you only a very tiny |
48 |
benefit (<5%). |
49 |
|
50 |
Such workloads are fairly unusual; as a gentoo user you will at |
51 |
minimum occasionally compile software, which almost always can be made |
52 |
parallel by the use of make's -j option. You might also consider |
53 |
that, for example, running openoffice and gimp at the same time also |
54 |
means X is running, and all 3 may well have work to do simultaneously. |
55 |
How well you exploit your system's CPU(s) depends on the very |
56 |
nitty-gritty details of the work you do. In general, SMP definitely |
57 |
does provide a noticeable performance boost in common situations. |
58 |
|
59 |
It would be nice to lay this issue to rest. The Ultra 5/10 and Blade |
60 |
1x0 are junk. They are similar in both design and performance to |
61 |
low-end PCs available at the times they were released. Other Ultra |
62 |
systems are significantly better and, although their CPUs are not |
63 |
clocked especially fast and their SCSI controllers are somewhat dated |
64 |
now, most offer both SMP and design characteristics which are only now |
65 |
becoming available in high-end PC workstations, such as switched, wide |
66 |
system interfaces. Finally, as always, understand your workload |
67 |
before buying. |
68 |
|
69 |
-- |
70 |
Keith M Wesolowski |
71 |
"Site launched. Many things not yet working." --Hector Urtubia |
72 |
|
73 |
-- |
74 |
gentoo-sparc@g.o mailing list |