Recommended CPU for timetabling

[Silver]

thank you.
i added all previous results in the fist message. (so also results from y).
thanks
since the first message grow and grow i just want to know this:
- is it needed to split it into 2 tables? one for windows and one for linux? (because the linux variants are always ~10% faster)
no, sort by the cpu better than sort by the system
- is it needed to cut results that are slower then maybe 5 minutes?
no, It's good publicity for Linux.

-------

[Silver]

1min 42s - Intel i7 - 2670Qm (2.2GHz - compiled and generated with Ubuntu 12.04 - 64 bit - metered by Silver)

[Liviu Lalescu]

The future version 5.18.0 will contain this bug fix: http://lalescu.ro/liviu/fet/forum/index.php?topic=1136 , and the random seeds and times will change from then on, sorry.

[Silver]

The future version 5.18.0 will contain this bug fix: http://lalescu.ro/liviu/fet/forum/index.php?topic=1136 , and the random seeds and times will change from then on, sorry.

i used fet 5.14 not 5.18 (from the Ubuntu repositories).

best regards

[yush]

I haven't compiled 5.14.3 on the Mac, so I used 5.15. Hope it's the same.

CPU: 2.4 GHz Intel Core i5 Dual Core Sandy-Bridge
Time: 2 min 27 sec

[Zsolt Udvari]

4min 5s - AMD Athlon(tm) X2 Dual-Core QL-65 (2100 MHz, compiled and generated on Linux)

Same machine, x86_64 build: 3 min 31 sec - 14% faster.

[Volker Dirr]

i guess it is faster because of an newer compiler and/or just because of "normal" inaccuracy (if you meter several times).
i don't think it is because of 64 bit, because there are not many 64 bit values in the source of fet and also 64 bit divide the cache in half (compared with 32 bit.) so 64 bit can be theoretical slower then 32 bit (of course only if the dataset use more then 50% and less then 100% of the cache in 32 bit mode.)
maybe you (or someone else) can proof that?

so maybe we also need to write down the used compiler?

[Zsolt Udvari]

I don't know - I'm not hardware/software expert - just now turned to 64bit my Arch and I was curious.
Do you think the newest gcc generates 14% faster binaries?

Edit: newest measurement: 3m39s (I've worked parallel too).

[Volker Dirr]

Difficult to say without more tests. But that is one possible reason. (Or you used other fet source in the past, or there where a few other programms running during the test in the past (in worst case a part of your operating system), or normal inaccuracy, or ...) Proof is only possible by several tests.

[Liviu Lalescu]

Maybe newer Qt. Maybe newer Linux (kernel).

64 bit should not be faster than 32 bits. Professor Knuth writes the same argument as Volker's on his page.

PS: My newer Qt is in fact slower in the interface. We may consider running next texts by command-line interface.

[Volker Dirr]

I added a new result:

1min 54s - Intel i5 3210m (2.5 GHz - precompiled FET with Windows 7  - metered by Volker Dirr)

Interesting to see that this new and expensive Intel cpu is not much better then an very old and cheap AMD cpu.

(2min 01s - AMD Athlon II X2 270 (3.40 GHz - compiled and generated with Linux - metered by Volker Dirr)

Is there someone that can test also on AMD E-450, Bulldozer, Trinity, Piledriver, ...? (By the way: A 4 core bulldozer will be faster then an 8 core bulldozer, because FET use only one core and the 4 core cpus have an higher clock speed.)

[Silver]

today, i updated Qt to 4.8 and i generated (german-100_and_0.fet) with seed x=1 y=1,

1min 02s - Intel i7 - 2670Qm (2.2GHz - compiled and generated with Ubuntu 12.04 - 32 bit - metered by Silver)

[Liviu Lalescu]

Is it 1 min 02 s or 1 min 20 s?

Volker will be unavailable for a couple of weeks, so he won't be able to update right away.

[Volker Dirr]

Hallo Silver,
maybe you used the latest FET version? For the benchmark only the old version is allowed.
So i will keep your old value:
1min 50s - Intel i7 - 2670Qm (2.2GHz - compiled and generated with Ubuntu 11.10 - metered by Silver)

[y]

3min 55s - Intel Core 2 Duo E6600 (2.40 GHz - precompiled FET 5.14.3 with Windows 7 32bit - metered by y)