1 |
Paul Hartman schrieb: |
2 |
> On Thu, Jun 11, 2009 at 3:07 PM, Kelly Hirai<kelly@×××××××.edu> wrote: |
3 |
>> the N270 is a single core with hyperthreading, which will apear as 2 |
4 |
>> cpus (with the same core id) in dmesg. |
5 |
> |
6 |
> Ah, I forgot about hyperthreading masquerading as multiple CPUs. In |
7 |
> that case, Maxim can safely disable SMP if he wants to. I don't know |
8 |
> if the theoretical speed gains of disabling SMP outweigh the |
9 |
> theoretical speed gains of enabling hyperthreading. I think it'll |
10 |
> probably be about the same either way. |
11 |
> |
12 |
|
13 |
Well, I don't know about real workloads but once I did a little |
14 |
benchmark: One versus two instances of `dd < /dev/zero | md5sum`. |
15 |
|
16 |
Two instances had a 30% higher throughput than one. I haven't tried it |
17 |
with disabled SMP but I really can't imagine that the extra scheduling |
18 |
would cost nearly enough to compensate for this. |
19 |
|
20 |
From a technical point of view, I think HT makes more sense for an Atom |
21 |
than for a Nehalem: The Atom has only one pipeline, no out-of-order |
22 |
execution and probably a less effective branch prediction. HT might |
23 |
compensate this. |
24 |
|
25 |
However, I'm wondering if it wouldn't have been better to implement |
26 |
out-of-order execution instead of HT (like VIA Nano, for example). Maybe |
27 |
HT doesn't need as many transistors as out-of-order execution? |
28 |
|
29 |
In the end, unless there is some hard evidence against the use of HT, |
30 |
I'd say: They've spend their transistor budget on HT, now we should use |
31 |
what we've got. |