Hello,

I have a question about the scores on the same machine after switching OS.

computer id=303222 mfps=2254.4 mis=4709.07
computer id=323181 mfps=1486.38 mis=2711.54

This is the same computer.
303222 = Windows i386 ( limited to 3.2 Gb RAM )
323181 = Linux x86_64 ( not limited )

The Linux machine uses less resources when running and still gets just above have the points. Why?

I do **not** think that the ia32 libs on Linux will slow down everything that much.

And yes we did install a SMP kernel.

With kind regards,
Marcel

---

Koru77's machine is showing off the difference between the Windows benchmarks and the linux benchmarks. That's why there were lots of complaints about Linux benchmarking under Boinc.

The good news is that the new credit system is giving Koru77 455.8 credits/day/core based on the very small 116,000 seconds worth of work. (small sample size.) Under Windows, Koru77 was getting 460.3 credits/day per core based on 674,000 seconds worth of work.

The Linux machine's scores should average out to the same it was getting under the Windows when there's a week's worth of WUs turned in.

---

I have some doubts about this.

I myself use 3 machines that run 24/7.

Computer ID=282024 Dual G5 PPC 2Ghz ( MacOSX )
Computer ID=304690 Core 2 Duo 6400 @ 2.13Ghz ( Linux x86_64 )
Computer ID=293595 Dual P3 500Mhz ( Linux i686 )

my RAC = +- 850

Now how could 303222 get a RAC of 750 at best?

see "http://stats.kwsn.net/user.php?proj=all&cpid=528d63d2398046e16f87f7a938450e26&period=4weeks"

Please don't tell me that the CPU speed ( 2.13Ghz -> 2.40Ghz ) has the same affect as an Dual G5 2Ghz.

Marcel

---

This is purely because the code generated in BOINC for Linux is "cr*p". Both Whetstone and Dhrystone are benchmarks that certain compilers will "obliterate" [I have, whilst experimemnting with compiler flags etc, managed to get a "score" of 36000MFlops on an Athlon64 at 2.2GHz - this is a machine that should in theory get 6600MFlops, if you can get EVERY SINGLE floating point unit to co-operate at the same time - but in this case I get about 6x that, because the compiler has removed most of the actual benchmark code and left it with a few empty loops]. Some compilers [with some settings] are better than others - and it's possible that some portion of the BOINC benchmark is not compiled "optimally" by the compiler for Linux - say for example that it's not set -ffast-math or some such, so the compiler HAS to call a function to calculate square-root, rather than actually performign the square-root instruction straight off.

[Doing square-root in -ffast-math is about twice as fast as doing it without -ffast-math].

There may be reasons why the compiler is set for some "not so great" performance on the benchmark.

For rosetta, the new credit system works reasonably well, so it's not a big problem.

And Marcel: You're right, the IA32 or x86_64 math, for this particular excercies is pretty much the same...


--
Mats

---

Is it fair to say that all the projects that use "special" credit calculation should be OS neutral? I'm thinking Rosetta, Einstein, SIMAP and SETI? Are there others?

The reason I ask is that I have Windows on all of my machines, but want to try Linux on one. I don't, however, want to trash my credit returns to the team.

---

Wave upon wave of demented avengers march cheerfully out of obscurity into the dream.

Is it fair to say that all the projects that use "special" credit calculation should be OS neutral? I'm thinking Rosetta, Einstein, SIMAP and SETI? Are there others?

The reason I ask is that I have Windows on all of my machines, but want to try Linux on one. I don't, however, want to trash my credit returns to the team.

I'd say that some credit calculations seem very "fair", others not so. Einstein is using some system similar to Rosetta, and Rosetta itself is "good" in the sense that it's not relying on the benchmark results directly. Any project where benchmark results are directly related to the credit given will be affected by the OS you're using - however, most projects are "averaging" somethihng or other, so if you run Linux, you'r claim will be lower, but it will average against someone else with a higher credit claim, and you'll end up with a higher score if you aren't really unlucky and get "all Linux systems" on the same work-unit - which is unlikely on Rosetta, as there's 80% Windows and about 10% Linux, with 10% of "others" thrown in for good measure.

--
Mats

---

I have a question about the scores on the same machine after switching OS.

computer id=303222 mfps=2254.4 mis=4709.07
computer id=323181 mfps=1486.38 mis=2711.54

This is the same computer.
303222 = Windows i386 ( limited to 3.2 Gb RAM )
323181 = Linux x86_64 ( not limited )

What version of linux and what processor? I found a similar problem when running ubuntu with a AMD or Core 2 Duo. Ubuntu intentionally throttles the performance for chips that allow it. The solution is to add "CPU Frequency Scaling Monitor" to the taskbar. Then you need to run a command (one time only) to give your userid permission to change the setting by running this in a terminal:

sudo dpkg-reconfigure gnome-applets

Now you can set your frequency to the maximum. However, I found that I needed to reset it to the max after every reboot.

You can read more here:

http://ubuntu.wordpress.com/2005/11/04/enabling-cpu-frequency-scaling/

---

Reno, NV
Team: SETI.USA

Any CPU frequency scaling governor worth it's salt should up the CPU speed to maximum quickly enough to not make any noticable difference over one minute. The sampling rate for the cpufreq driver is in the "seconds" range even for the conservative model of cpu frequency changing.

Further, I have poor cpu benchmarks on all of my machines, and none of them run powernow drivers...

--
Mats

---

Is it fair to say that all the projects that use "special" credit calculation should be OS neutral? I'm thinking Rosetta, Einstein, SIMAP and SETI? Are there others?

The reason I ask is that I have Windows on all of my machines, but want to try Linux on one. I don't, however, want to trash my credit returns to the team.

Yes and No. :-)

Each scientific application needs to be compiled under that platforms OS, therefore it needs to be compiled for MacOS-X, Linux and Windows. Each OS has a different compiler and may be better or worse than the one under another OS. So linux could compile generate code that its slower then under Windows which may be slower than under Mac. If you run it on the same hardware.
Of course there is also OS overhead, one OS may need more time just do do it's own thing.

---

Team mauisun.org

Any CPU frequency scaling governor worth it's salt should up the CPU speed to maximum quickly enough to not make any noticable difference over one minute. The sampling rate for the cpufreq driver is in the "seconds" range even for the conservative model of cpu frequency changing.

I can speak only for my own experience. I was getting *very* poor performance, about 50% of what it should have been. When I manually set the speed of the processors, it fixed the problem. So I am suggesting that Marcel try my solution (assuming he is using an AMD or Core (2) processor).

You can read about my problem and solution here:

https://boinc.bakerlab.org/rosetta/forum_thread.php?id=2218

I am guessing that perhaps because of the low priority of the BOINC processes, ubuntu did not think it was required to automatically increase the clock speed.

---

Reno, NV
Team: SETI.USA

Zombie67.

We run Debian "etch" on Core 2 Duo ( me 3.16Ghz koru77 on 3.40Ghz ).
Basically we are running VMware Server on this machine.
No GUI is installed.

We both have 4Gb RAM for this we need the EM64T part.
We run BIONC under the ai32 libs.

With kind regards,
Marcel

---

Ok, I'm talking about the fact that one machine (running windows) gets:
Measured floating point speed 2051.84 million ops/sec
Measured integer speed 3794.71 million ops/sec

Details: https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=309817

And onother (running Linux):
Measured floating point speed 1054.06 million ops/sec
Measured integer speed 2005.7 million ops/sec

Details: https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=306057

But they both have Athlon64 X2 4400+ processors - ok, so the Windows machine is only using one core (because the IT department gives us a standard windows installation with only single processor support) - but the benchmark scores are PER CORE, so it should have the same score if the benchmarks were the same.

cat /proc/devices/system/cpu/cpu0/cpufreq/cpuinfo_cur_freq says "2200000", so 2200MHz. This machine spends 99% of it's time running Rosetta or some other BOINC project.

--
Mats

---

Marcel:
I showed that your teammate's system was getting the same credit for work done under windows as they are under linux.
They get about 460 credits per core per 86400 seconds of work. One core, running 24/7 should turn in near 604,000 seconds of work. A dual core system should turn in around 1,208,000 seconds of work.
Your teammate turned in around 800,000 seconds of work for a week.

Your machine is getting 347 credits per core per 86400 seconds of work. You've turned in 920,000 seconds of work for the week. For some reason, you're missing 288,000 seconds of work for the week. Any idea what's eating up that much cpu time?

So Koru77 isn't getting the full 1,208,000 seconds of work a week that the machine is capable of; and as Zombie stated, your cpu may be throttled since it should be producing closer to the 460 credit/core per 86400 seconds of work of Koru77's 2.4 Ghz 6600; since your 6400 is running at 2.13.

The frequency ratio is 1.12 (2400/2130), the benchmark ratio is 1.17, but the credit/second ratio is 1.31. Does the 6600 core 2 duo have twice the cache of the 6400 that would allow it to perform better than the frequency jump would suggest?

---

Zombie67.

We run Debian "etch" on Core 2 Duo ( me 3.16Ghz koru77 on 3.40Ghz ).
Basically we are running VMware Server on this machine.
No GUI is installed.

I believe that ubuntu is debian. I think you may be having the problem I described. However, I do not have experience to manually set this via command line. The link I gave earlier may lead you in the right direction. This is the part that applies here, I think:

---

In order to be able to change the operating frequency, your processor should support changing it. You can find out if your processor has scaling support by seeing the contents of files in the /sys/devices/system/cpu/cpu0/cpufreq/

For example, on my system:

$cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_frequencies
gives:

1300000 1200000 1000000 800000 600000

Which means that the above frequencies (in Hz) are supported by my CPU.
and…
$cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_governors
gives:

userspace powersave ondemand conservative performance

All those are the different “modes” I can operate the CPU at. Userspace, for example, regulates the frequency according to demand. Performance runs the CPU at max-frequency, etc…

On the Ubuntu Forums, I read that one can manually change the frequency by executing commands like:
$ cpufreq-selector -f 1300000
which will set the frequency to 1.3 GHz.

---

Good luck

---

Reno, NV
Team: SETI.USA