Do you have to leave Archlinux for the AMD bulldozer

Pi · 2011-11-25 03:20:55

It's not a bulldozer, but for curiosity sake I just tried the Arch kernel compile on my stock i7-970. Went through and used abs with a -j 12 in makepkg.conf. Actually forgot that it had to download the source first so this figure includes everything from start to finish. I'm thinking it took about 30 sec to a minute or so to download the 61 meg <shrug>. But anyway here's what I got

==> Finished making: linux 3.1.1-1 (Thu Nov 24 19:03:28 PST 2011)

real	10m4.558s
user	73m42.488s
sys	6m43.434s

.:B:. · 2011-11-25 07:01:51

I wonder if my i3 is that powerful then (it has HyperThreading, but that's nothing like 4 physical cores). Compiling Android didn't take that long - at least it felt like that. I didn't exactly sit around waiting .

Bulldozer seems to do very well in multithreaded environments, but it seems you actually loose performance when it comes to singlethreaded. Of course, if multithreaded usage fits your profile... I don't know what to think of it myself, of AMD's new breed.

korkadapa · 2011-11-25 09:38:20

I forgot to use all four cores. It's down to 17m35.916s on the Q6600 now. But still, it's quite a lot slower than the i7s.

Earnestly · 2011-11-25 11:30:16

Lone_Wolf wrote:

It might have to do with the compiler used for the benchmarks.
http://www.agner.org/optimize/blog/read.php?i=49#49

This is very interesting as I'm thinking of upgrading to a new machine which I can build upon for the next few years to come.

Does this, combined with the results from Vamp868, conclude the Bulldozer to be a better CPU in so far as other benchmarks are "rigged"? And what are the implications of such an underhanded technique by Intel? But more importantly, is Intel still doing this?

( Forgive me, but I'm still fairly new to PCs in general after many years of Windows usage I know almost nothing about the actual components beyond numbers. )

Thanks in advance.

krum · 2011-11-25 12:06:45

Vamp898 wrote:

Vor sure i disabled CPU-Freq completely or set to "performance" before i run the Benchmarks.
I also run every Benchmarks about 5-6 Times and the changes was about 1-2% from run-to-run
[...]
The Bulldozer was faster on really everything. So what if you think the i7 2600 is faster on something, tell me on what, i will test it.

Could you please provide some logs and numbers? (for comparison and so on)
Thx!

sakisds · 2011-11-25 12:27:57

korkadapa wrote:

Yes, I'm aware of ccache. Compiling large projects is a pain anyways, especially android takes forever to compile. That's why I'm interested in the bulldozer, it seems to offer great multithread performance while still having a lower price than the 2600k. But if I upgrade I would like to see a great speed increase, if the increase will be barely noticeable I might just as well wait a year or two for my upgrade. It seems though, from my simple test here, that the latest generation CPUs offer a great speed increase, given that sakisds didn't use stuff like ccach.

Nop, I didn't use anything special. I just downloaded the source and used 'make -j5'. I don't compile nearly enough to bother but I might look at ccache sometime.

Vamp898 · 2011-11-25 22:20:55

First of all, its increadible how long the Arch Kernel seems to compile. I always configure and compile my kernel and it takes less than two minutes to compile xD i tihk i give the stock archlinux kernel a try

@krum I used a LibreOffice Calc Sheet where i noted down all results and i can provide you that. That is what i used to create these fancy 3D Charts (i really like this 3D Effect xD).

I can also run some benchmarks again if you like (but i need someone which takes the i7 part for me, i already sent the i7 back. It was over 340 € + mainboard and i dont want to waste that money )

Anyway i did those benchmarks on Gentoo with all optimisations on (especially -march=native) and it seems like the speed difference to Arch is sometimes there, sometimes not.

But i think a quite good Benchmark is to download the default Linux Kernel from here: http://www.kernel.org/pub/linux/kernel/ … .2.tar.bz2 run make defconfig and than "time make -jwhatever"

that should give a comparable thing (even thorough the used GCC version and so on maybe change the result)

Vamp898 · 2011-11-25 22:39:46

So here is my result with Linux Kernel 3.1.0 (3.1.2 tooked to long to download at 400kb/s so i just used my 3.1.0 copy, i hope that wont have any significant speed change. Maybe you also want to download 3.1.0).

This is my result: http://img171.imageshack.us/img171/2314/bulldozer1.png

_without_ make defconfig, so this is only the make itself, but the whole make.

So if you also want to post your result please to the follwing

take a _fresh_ extraction of the linux kernel, run _only_ make defconfig and directly after that "time make all -j12" (or -j5 or whatever the amount of cores + (cores/2) gives for you.

btw. i also run this thing on Gentoo before (did make clean between those) and i was a bit disappointed that the fully optimized Gentoo only was about 0,5 seconds faster xD so it seems like optimisation dont increase compilation speed

btw. does anyone knows why the "user" time says 12m? it never tooked 12m for sure, i watched on the clock and this is just impossible. the "real" colum exactly says the time it used to compile i think i didn´t got the "time" command completely. I thought user was the time he spend in user mode and user + sys = real, but it seems kinda different

Last edited by Vamp898 (2011-11-25 22:43:11)

krum · 2011-11-25 23:07:22

Maybe user time is related to (time x number of core) or something like that...

korkadapa · 2011-11-25 23:11:42

I did the same with my Q6600. Linux 3.1, make defconfig and make -j6.

real	3m30.665s
user	8m22.711s
sys	1m10.985s

So quite a lot slower than the bulldozer. I hope that someone with an i7 would post their result too. Preferably a Sandy bridge.

Pierre · 2011-11-26 00:35:25

I'd say it's hard to compare with quite different systems and setups. I ran it for fun on my i7 2600k and my result was quite similar: "make all -j8 495,59s user 39,67s system 694% cpu 1:17,03 total" As you see I just used j=core*2 here.

Vamp898 · 2011-11-26 09:09:53

cores*2?

so that means you dont use HT?

a typical linux system is quite well multithreaded and i seen an speed increase on most stuff about 15-20% with HT enabled. Is there a specific reason why you dont use HT?

Seems like something else than the processor is slowing down my system. Even with HT enabled i only got about the same result as with my bulldozer when compiling the linux kernel.

Pierre · 2011-11-26 09:21:33

No, I use HT, but my i7 only has four cores. So I compiled with 8 processes.

Vamp898 · 2011-11-26 10:01:36

Hmm so GCC gives every logical core one process, maybe there could still exist gaps but im not sure about that. I dont know if the scheduler considers the HT or just random gives the processes to the logical cores.

sakisds · 2011-11-26 10:14:44

Vamp898 wrote:

So here is my result with Linux Kernel 3.1.0 (3.1.2 tooked to long to download at 400kb/s so i just used my 3.1.0 copy, i hope that wont have any significant speed change. Maybe you also want to download 3.1.0).
This is my result: http://img171.imageshack.us/img171/2314/bulldozer1.png
_without_ make defconfig, so this is only the make itself, but the whole make.
So if you also want to post your result please to the follwing
take a _fresh_ extraction of the linux kernel, run _only_ make defconfig and directly after that "time make all -j12" (or -j5 or whatever the amount of cores + (cores/2) gives for you.
btw. i also run this thing on Gentoo before (did make clean between those) and i was a bit disappointed that the fully optimized Gentoo only was about 0,5 seconds faster xD so it seems like optimisation dont increase compilation speed
btw. does anyone knows why the "user" time says 12m? it never tooked 12m for sure, i watched on the clock and this is just impossible. the "real" colum exactly says the time it used to compile i think i didn´t got the "time" command completely. I thought user was the time he spend in user mode and user + sys = real, but it seems kinda different

Ok then, I will recompile the kernel fresh from the tar on my 2500. Last time I copied arch's default copy (like it says at the wiki). I will post the results in a while.

Edit: With make defconfig; time make -j5 I got this:

real	1m36.593s
user	0m0.240s
sys	0m0.020s

Last edited by sakisds (2011-11-26 10:27:11)

Vamp898 · 2011-11-26 12:35:13

That really seems like something is breaking down my system. Weven with the i7 2600 i didn´t got this times.

Also i have sometimes random segfaults with my older Phenom which seems like something is wrong with my memory...

I try to plug some other memory or maybe just remove 2 of my 4 memory slots (i think the two new one aren´t the bests) and run it again.

in theory the i5 2500 should be slower than the i7 2600, but Pierre too had a much better time than me with my i7.

I hope it is the memory xD otherwise i really dont know whats going on

Hmmm ok its not the memory, but while testing this i got some other ideas lemme check.

bingo,

i dont know why but he wrote into the swap because my memory was filled with other stuff. I dont know why, i still had 5GB of free memory (of 16GB) but anyway he started to write into the swap while compiling.

a "swapoff /dev/sdb2" and babam, now it looks different

Kernel: arch/x86/boot/bzImage is ready (#9)
real 0m56.753s

but i used make -j512 now to compile it xDD but it really doesnt seem to be faster than with -j12 or -j16, but if you persist on the jobs i will re-compile it with -j12

Also you can try with -j512 and look if it maybe speed up but i really dont think so, 100% cpu is 100% cpu

also i used a tmpfs which seems to increase the speed by about 10seconds, again, if you persist i will run again without tmpfs

//Edit: Without a tmpfs i get about 59 Seconds, so not fully 10 seoncs but a speedup anyway. Specially at the end of compiling there seems some speedups when hes going to copy stuff around and gzip it.

btw. with the boxed cooler the maximum temp i get is about 45-50°, mostly hes at 30-40° (when all cores are at 100%). And when he is finished, he goes down to 20° in about a half minute. Right now hes on 16°. I heared those Intel processors are harder to cool even when they need less power. Either they have bad boxed coolers ore they seem to get hotter

Last edited by Vamp898 (2011-11-26 13:23:44)

Lone_Wolf · 2011-11-27 13:02:27

Kaustic wrote:

Lone_Wolf wrote:
It might have to do with the compiler used for the benchmarks.
http://www.agner.org/optimize/blog/read.php?i=49#49
This is very interesting as I'm thinking of upgrading to a new machine which I can build upon for the next few years to come.
Does this, combined with the results from Vamp868, conclude the Bulldozer to be a better CPU in so far as other benchmarks are "rigged"? And what are the implications of such an underhanded technique by Intel? But more importantly, is Intel still doing this?
( Forgive me, but I'm still fairly new to PCs in general after many years of Windows usage I know almost nothing about the actual components beyond numbers. )
Thanks in advance.

it does mean that only benchmarks compiled with gcc can be fully trusted, unfortunately there are not many of those.
While phoronix does a lot of linux hardware tests, it's test are often flawed or use wrong settings.

Implications :
Intel has used unfair techniques to make competitors look bad.
Since the terms of the settlement between amd and intel are mostly secret, we don't know if and when intel has changed their compiler behaviour

Based on my extensive experience with different hardware systems whether intel or amd processors are better for you depends largely on 2 things :
- budget
For processors below approx 150 USD Amd tends to give more bang-for-your buck

For segment between 150 and 300 USD :
Singlethreaded mainly : Intel
multi-threaded mainly : AMD

Top segment : Intel

NOTE that this is valid for consumer processors.

For server systems :
for single processor systems the most important factor is singlethreaded performance (intel) vs multithreaded performance(amd).
for servers with 2 or more processors AMD is usually the better choice.

Earnestly · 2011-11-27 13:34:30

Thanks for your time and explanation, it cleared a few things up.

I'd like to try support the least corrupt companies if possible ( of which I knew nothing about when using Windows ) and as I don't need an incredible CPU I might now move to AMD in my next build, possible ATi as well. At least their cpu sockets last longer as well.

Vamp898 · 2011-11-27 15:16:17

I, for sure, compiled hardinfo by hand using GCC to run the benchmarks and as you can see AMD wins by far.

Maybe i will give AMDs Open64 Compiler a chance which should increase the speed von the AMD even more.

About corruption, AMD seems to be less corrupt. At least they didn´t had that much problems with the curt also AMD is very friendly to the opensource world. After they bought ATI the provieded all documentations. They provide a lot of documentations how to optimize your processor using GCC and AMDs own Open64 (which is just a variant of GCC).

Also AMD have better prices.

For 1000 € on Intel you get the Core i7-3960X with 6x3,3 GHz and about 16mb of Cache (every core 256kb + 15mb of l3).
For 1000 € on AMD you get the Opteron 6282 SE with 16x2,6 GHz and 32mb of Cache (every core 1mb + 16mb of l3)

i think the extreme edition processors from Intel are just bad jokes to get even more money with people which are that insane to pay 1000 € for a processor which value is maximum 500 €.

I would really like to Benchmark the Core i7-3960X with a real optimised opensource benchmark.

The Problem with most Benchmarks is

1. They are mostly compiled without any optimisation (-O0 which never happens in real world)
2. The Windows one seems to be all closed source. No one knows the source, how the benchmark works, what it does and if its true

And that is what i wonder about.

AMD looses on all those benchmarks, but if you take 1 optimized opensource benchmark the seem to have _much_ better cards (and every system today uses minimum -O2. The Linux Kernel even can´t compile with -O0 (you get some compiler errors))

But the result is still very real, intel have a eye on raytracing and so even on optimized benchmarks, AMD loose by far on Raytracing. But they win on every other benchmark anyway.

So in the end the Benchmarks are opensource, you can read every single line to check, my benchmark was optimised (for the intel and for the amd too) and the result seems to be very real.

And that is why i kept the Bulldozer and send the intel back. They have a more corrupt idea of leading a company, insane prises in the high-end segment and they seem only to win if the benchmark is closed-source and not-optimized (which is a really strange)

Even the Phenom II X6 1100T won in all benchmarks (except raytracing) against the intel core i7 2600. As i said, on the Windows Benchmarks the Phenom II X6 1100T is on place 16 or even worse behind nearly all intels.

That is just absurd.,

//Edit:

Just think back to the days where AMD Introduced 3DNow which was much faster than Intels SSE, but developer do not used 3DNow and used SSE and so 3DNow, even when it was faster/better, was a fail.

That shows how Intel is focused on Money and even derail development to get more money.

Who knows how far technice and development would be today if Microsoft and Intel wouldn´t rule the market.

//Edit2:

btw. there are some key-aspects in which you can see that my Benchmarks can not be that wrong. Even when intel is slower than amd on all benchmarks, he is much faster on the Raytracing. In case that intel have a special hand on raytracing that explains why he is faster. If the Benchmark would be wrong at all, intel would be randomly faster/slower or at all faster/slower but not only on one thing where he is very optimised.

Also the Bulldozer is on FPU Benchmarks slower than the Phenom, but mostly faster on the others. The Bulldozer have only 4 FPU Units and the Phenom 6 FPU Units which explains why the Phenom is faster. Everything else than random, more likely good to explain and to understand why it is correct how it is.

So there is maybe a toleranz on the benchmark but the most results are very good to explain and traceable than beeing just random or wrong.

The only thing i really dont understand and be able to explain is the N-Queens benchmark. I really dont got why the bulldozer is that fast here, but that is positive for the intel so that is most likely not the benchmark where intel fans would say "Hey this is all wrong". But in the end i still dont got why N-Queens is so slow on Bulldozer. This is 100% reproducable even when running the test 100 times.

Last edited by Vamp898 (2011-11-27 16:59:52)

Lone_Wolf · 2011-11-27 19:29:02

Vamp898 wrote:

The only thing i really dont understand and be able to explain is the N-Queens benchmark. I really dont got why the bulldozer is that fast here, but that is positive for the intel so that is most likely not the benchmark where intel fans would say "Hey this is all wrong". But in the end i still dont got why N-Queens is so slow on Bulldozer. This is 100% reproducable even when running the test 100 times.

This is likely due to the differences in the processor architecture.
Intel procs tend to have a long instruction pipeline, and lean heavily on branch prediction to reduce the number of times the instruction pipeline needs to be discarded/reloaded.

Amd procs typically have a shorter pipeline that takes less time to reload, but a much worse branch prediction.

Another example of an algorithm which performance is heavily influenced by proc architecture is the Prime algortihm.
In the Intel P4/Amd Athlon 64 era and had a much faster memory access method then intel.
To compensate for intel's slow memory access intel put 2-4 times as much cache memory in their procs as Amd did.

On many intel procs prime could run completely in cache memory, while on amd procs prime did not fit in cache memory and needed to use main memory.
This made prime useless for comparing performance.

Vamp898 · 2011-11-28 08:17:01

Good to know thx

Vamp898 · 2011-11-30 17:59:54

If people are still willing to compare processor performances. I started to write my own Benchmark Suite today and some self-wrote Benchmarks are already finished.

If you want you can download and run it here (source + compile skript): http://www.ignaz.org/ignaz_bench.tar.gz

You can post your results here

Earnestly · 2011-11-30 22:38:46

This is great, I'd really like to see more comparisons on an open platform to other powerful CPUs, if you can contribute that'd be wonderful.
Thanks

karol · 2011-11-30 22:48:47

Lone_Wolf wrote:

Based on my extensive experience with different hardware systems whether intel or amd processors are better for you depends largely on 2 things :
- budget
For processors below approx 150 USD Amd tends to give more bang-for-your buck
For segment between 150 and 300 USD :
Singlethreaded mainly : Intel
multi-threaded mainly : AMD
Top segment : Intel

http://www.tomshardware.com/news/AMD-AP … 14114.html

paziul · 2011-12-01 02:58:11

probably many or most of you are familiar with ars technica's site, and none of the information sources should be considered as "the only and the best", generally they are often not too far from being neutral and unbiased, have a quite "healthy" point of view in many cases - also, often polemize - not this time....:

bulldozer @ arstechnica

BTW: IgnazBench on 4GHz i5 650
1: 5.520325
2: 7.170555
3: 9.715491

Arch Linux

#26 2011-11-25 03:20:55

Re: Do you have to leave Archlinux for the AMD bulldozer

#27 2011-11-25 07:01:51

Re: Do you have to leave Archlinux for the AMD bulldozer

#28 2011-11-25 09:38:20

Re: Do you have to leave Archlinux for the AMD bulldozer

#29 2011-11-25 11:30:16

Re: Do you have to leave Archlinux for the AMD bulldozer

#30 2011-11-25 12:06:45

Re: Do you have to leave Archlinux for the AMD bulldozer

#31 2011-11-25 12:27:57

Re: Do you have to leave Archlinux for the AMD bulldozer

#32 2011-11-25 22:20:55

Re: Do you have to leave Archlinux for the AMD bulldozer

#33 2011-11-25 22:39:46

Re: Do you have to leave Archlinux for the AMD bulldozer

#34 2011-11-25 23:07:22

Re: Do you have to leave Archlinux for the AMD bulldozer

#35 2011-11-25 23:11:42

Re: Do you have to leave Archlinux for the AMD bulldozer

#36 2011-11-26 00:35:25

Re: Do you have to leave Archlinux for the AMD bulldozer

#37 2011-11-26 09:09:53

Re: Do you have to leave Archlinux for the AMD bulldozer

#38 2011-11-26 09:21:33

Re: Do you have to leave Archlinux for the AMD bulldozer

#39 2011-11-26 10:01:36

Re: Do you have to leave Archlinux for the AMD bulldozer

#40 2011-11-26 10:14:44

Re: Do you have to leave Archlinux for the AMD bulldozer

#41 2011-11-26 12:35:13

Re: Do you have to leave Archlinux for the AMD bulldozer

#42 2011-11-27 13:02:27

Re: Do you have to leave Archlinux for the AMD bulldozer

#43 2011-11-27 13:34:30

Re: Do you have to leave Archlinux for the AMD bulldozer

#44 2011-11-27 15:16:17

Re: Do you have to leave Archlinux for the AMD bulldozer

#45 2011-11-27 19:29:02

Re: Do you have to leave Archlinux for the AMD bulldozer

#46 2011-11-28 08:17:01

Re: Do you have to leave Archlinux for the AMD bulldozer

#47 2011-11-30 17:59:54

Re: Do you have to leave Archlinux for the AMD bulldozer

#48 2011-11-30 22:38:46

Re: Do you have to leave Archlinux for the AMD bulldozer

#49 2011-11-30 22:48:47

Re: Do you have to leave Archlinux for the AMD bulldozer

#50 2011-12-01 02:58:11

Re: Do you have to leave Archlinux for the AMD bulldozer

Board footer