Linux versus Windows - bigadv folding results

Axdrenalin,
I guess this is what happens when i post in the wee hours of the morning
I do indeed have the 6.34 client and passkey. Currently stuck with an annoying P6702 :(

Axdrenalin said:
Welcome RETARDOCAU. nice to have you on the forums. You might want to make sure that you're using the latest client (6.34 I think) and are using a passkey as well...I didn't see that mentioned in your post.
Click to expand...

may i be worthy,
I'm sure we'd love you on T24 but im an ethical guy and if your happy amongst [H] then thats great news :) Perhaps you could use my motto instead of my name.
"I'm not smart but i can lift heavy things" :p

as for the lack of BA units - i have read the thread on the FF. Its frustrating to see people around you getting linux BA on their first try lol. I'm can be patient but was more concerned on whether i had missed something in my set up.
I've read a lack of allocated ram also using -advmethods can result in no BA.

As a guide you would expect that my smp ppd is approx half what my -bigadv could be?

may i be worthy said:
Hey RETARDOCAU, I am the lone Aussie on [H] here... welcome. I could borrow your username, because when I started folding I didn't even realise ocau had a team. (don't hate me). I spent one naive day with EVGA (don't hate me #2) before moving over here, where it is nice and warm.

I think there are server issues with giving out bigadv on linux atm. see here.

I have been testing linux builds for a week, and now that I have settled on musky's mix as my best so far, I have started folding on my 2 slaves. The one that has had the HDD/install that did all the testing has yet to get a bigadv in about 2 days folding over the last few days. The fresh install I did last night on my 3rd rig scored a bigadv 6901 out of the box. Between units just now on my 2nd box I just did a fresh install of FAH, but no, still cursed, drew a SMP. On non-6701 units I am seeing exactly half the PPD I see on 6901s. - on windows it is more like 2/3.

Interestingly the SMP frame times (still Core A3) on linux seem to be within 2 seconds of the times they always were on windows, whereas Core A5 bigadv is running 11-12% faster = 19-20% better PPD. :D

My main rig is my workstation and has to stay windows, but all I want to do is see all three pumping away at max before I have to shut down tomorrow. Might not happen.

My HFM is in my sig.
Click to expand...
 
I have being having ppd in the trash since the client switch... and already made sure that they all were on the new client..

someone suggested that I check and make sure the passkey stuck... you might want to try that...
 
thanks Patriot

im getting the bonus points with out issue but all my units are on the A3 core instead of A5.
 
RETARDOCAU said:
as for the lack of BA units - i have read the thread on the FF. Its frustrating to see people around you getting linux BA on their first try lol. I'm can be patient but was more concerned on whether i had missed something in my set up.
I've read a lack of allocated ram also using -advmethods can result in no BA.

As a guide you would expect that my smp ppd is approx half what my -bigadv could be?
Click to expand...

Unfortunately, Linux bigadv units do appear to be in short supply right now. Of the last four of my machines that finished a bigadv unit, two got another bigadv and two got a regular SMP unit. It is really hit and mis right now. I have run 5 or so regular SMP units in a row on a couple of my machines. i doubt your setup is wrong.

As far as ppd on regular versus bigadv, 50% is a good guideline. 6701/6702s are probably 1/3, and 71xx units are a little worse than half. The rest are really close to half.
 
I am sure you will get one soon retardocau - I did eventually get a bigadv on #2, so it was never a config issue.

On my 2 new linux installs I have been on bigadv for 2/3 of the time this week. - 2 bigadv, 10 SMP - one 6701. As Linux = 20% better PPD on bigadv, and 50% or worse on SMP, my break even point is doing bigadv 70% of the time or better.

Hmmm, but that does not account for 2684s, which after all just as bad as SMP for 24 hours straight... argh - too late for maths, but my gut feeling is that if the 6900/6901s are still flowing heavily I suspect it could be better to stick booted to Windows until they fix this.

Having said that, while testing Linux out for the first time it has been ok to have some SMP to start with while working the gremlins out.
 
musky said:
Unfortunately, Linux bigadv units do appear to be in short supply right now. Of the last four of my machines that finished a bigadv unit, two got another bigadv and two got a regular SMP unit. It is really hit and mis right now. I have run 5 or so regular SMP units in a row on a couple of my machines. i doubt your setup is wrong.

As far as ppd on regular versus bigadv, 50% is a good guideline. 6701/6702s are probably 1/3, and 71xx units are a little worse than half. The rest are really close to half.
Click to expand...

My SR-2 hasn't seen a BigAdv in a couple of days. Not surprising since p6901 seems to be the only unit running on the Linux side of things. This happens and usually will get resolved in a couple of weeks (fingers crossed) at the latest.

Funny enough the 71xx units fly on the AMD to a tune of nearly 180K PPD vs. a lot less on the SR-2. The SR-2 gets 90K on the 60xx units and less than 85K on the 71xx due to a difference in TPF of about 4 seconds..

The AMD is back on Windows for today. Will try and fix the Gentoo install on it tonight OR just throw a PCI ethernet card into it temporarily to test.
 
may i be worthy said:
I am sure you will get one soon retardocau - I did eventually get a bigadv on #2, so it was never a config issue.

On my 2 new linux installs I have been on bigadv for 2/3 of the time this week. - 2 bigadv, 10 SMP - one 6701. As Linux = 20% better PPD on bigadv, and 50% or worse on SMP, my break even point is doing bigadv 70% of the time or better.

Hmmm, but that does not account for 2684s, which after all just as bad as SMP for 24 hours straight... argh - too late for maths, but my gut feeling is that if the 6900/6901s are still flowing heavily I suspect it could be better to stick booted to Windows until they fix this.

Having said that, while testing Linux out for the first time it has been ok to have some SMP to start with while working the gremlins out.
Click to expand...

I am actually curious if they have a graded system when they are short ... my 24core has never been without... my 8 core system has frequently...

on the beta team we have a few more wu open maybe thats why I have not seen the 6901 in a bit...
 
Patriot said:
I am actually curious if they have a graded system when they are short ... my 24core has never been without... my 8 core system has frequently...

on the beta team we have a few more wu open maybe thats why I have not seen the 6901 in a bit...
Click to expand...

I would confidently say that they do not have any rhyme or reason to what machines get what units when there is a shortage. If that were the case, my SR-2s would never be without bigadv units. As it is, my slowest bigadv box has run nothing but bigadv units since I switched it over to Linux and my SR-2s are starving for them.
 
It's confirmed. BFS does not scale well to high core counts. 24 seems to be the limit.

On the 48-core box, using ZEN Kernel with BFS has elevated frame times by 50% over Ubuntu running CFS.

I saw similar slow downs with Ubuntu and the CK patches to add BFS. S field over at EVGA confirmed this as well. He believes, based on what he's read that this is the case.

Alternatively, my SR-2 running Gentoo with BFS is running at 10:55 minutes TPF after finally (!) receiving a p6901 last night for 170K PPD. The most it ever did in Windowz was 145K with 12:00 minute frame times.
 
Yeah the author of BFS didn't think it would scale well past 16 cores. THE FAQ isn't exactly clear, since all the 0.300 changes tend to be tacked onto the end of each section of the FAQ, but it does say
It is NOT NUMA aware in the sense that it does any fancy shit on NUMA, but
it will work on NUMA hardware just fine. Only the really big NUMA hardware
is likely to suffer in performance, and this is theoretically only, since
no one has that sort of hardware to prove it to me, but it seems almost
certain. v0.300 onwards have NUMA enhancements.
Click to expand...
 
Vediovus said:
Yeah the author of BFS didn't think it would scale well past 16 cores. THE FAQ isn't exactly clear, since all the 0.300 changes tend to be tacked onto the end of each section of the FAQ, but it does say
Click to expand...

I guess now people have hardware to prove it to him :D

There's three quad-AMD nutjobs on the FoldingForum who (I believe) own them, and one who has access to it.

I plan to re-compile the Zen Kernel with CFS and see what happens. I've also compiled the Gentoo-sources (Gentoo) kernel as another entry in my grub.conf file. I have five entries in it now. :eek:

It seems Zen Kernel needs explicit definition of IOV under PCI options to enable the Intel 82576 drivers. I've been looking for this option for a few days now. So many options ugh.
 
well I just got my folding farm back working.... don't feel like tweaking it just yet...
 
I know the feeling. I've got a lot of tweaking I've been putting off. I need to find some 80MM PWM fans for the heatsinks on my C32 machine that aren't loud as hell, get a new ups, fix the sata power connector on the psu I managed to break, overclock the desktop some more, find a modular psu for the htpc so I can put the side panels back on it, wait for ups to show up with my new hdd for the desktop, and test a mobo to see if I can get a little more ppd going. Thats just the farm projects, I have a car in the garage I need to see if its worth fixing too. I guess its a good thing I'm off til Tuesday :rolleyes:
 
Quick update:

Finally got the ethernet problem fixed on Gentoo with the AMD. The module refused to load under Zen Kernel even though it is a 2.6 module but whatever. Genkernel is on now.

Even compiling with regular "make" caused the igbvf modules to NOT load for the ethernet (Intel 82576) so I had to GenKernelize it. Zen is out.

It seems that CFS is the best way to go. TPF on a 6065 is back down to 0:49 with NUMA disabled. This is the equivalent I had with straight Ubuntu (without BFS patches).

This is strange because the opposite happened on Ubuntu (earlier) where NUMA was flying high. Either way I'll watch it, and hopefully get a BigAdv on this bastard this weekend.
 
^^ Critical stuff like NIC drivers and such, I compile directly into the kernel. Modules are nice for somethings but not all.
 
Ha! Ethernet drivers. I forgot those the first two times I compiled my zen kernel. The first time because I forgot ext4 support with / being ext4 :rolleyes:
 
The sad thing about linux bigadv A5 getting better PPD, is that now the difference between best and worst units points grows even wider.

These machines are 2% apart in speed, but 350% apart in points: :eek:

original.jpg
 
smp gets a low k value... 6701 4.1 vs 26.4 and smp has never scaled well past 6 cores really lol...
 
may i be worthy said:
The sad thing about linux bigadv A5 getting better PPD, is that now the difference between best and worst units points grows even wider.

These machines are 2% apart in speed, but 350% apart in points: :eek:
Click to expand...

There's a 220K PPD point differential on the AMD so I can relate, empathize, and co-miserate.

Thank dog the SR-2 has a regular SMP unit, and not one of "those"

Patriot said:
smp gets a low k value... 6701 4.1 vs 26.4 and smp has never scaled well past 6 cores really lol...
Click to expand...

They scale fine, just not as well as BigAdv. With the AMD going from 24 to 48-cores they went pretty much half TPF, so performance-wise, they scale just fine.


Tobit said:
^^ Critical stuff like NIC drivers and such, I compile directly into the kernel. Modules are nice for somethings but not all.
Click to expand...


I get really, really strange variances with Make menuconfig. It all seems to be related to what the current, loaded configuration sees. Even the original boot with the Minimal Install CD gave me a kernel generated where I didn't even SEE any gigabit ethernet drivers as options.

Now it's complaining about a missing IPV6 module. Lol. At least it boots, but doesn't seem any faster than Gen Kernel.
 
10e said:
They scale fine, just not as well as BigAdv. With the AMD going from 24 to 48-cores they went pretty much half TPF, so performance-wise, they scale just fine.
Click to expand...

huh... not the same experience I had... 1 smp was getting me ~13k on my 24core @ 1.7ghz

dual smp -12 were running about 10k a piece ...

10e said:
I get really, really strange variances with Make menuconfig. It all seems to be related to what the current, loaded configuration sees. Even the original boot with the Minimal Install CD gave me a kernel generated where I didn't even SEE any gigabit ethernet drivers as options.

Now it's complaining about a missing IPV6 module. Lol. At least it boots, but doesn't seem any faster than Gen Kernel.
Reply With Quote
Click to expand...

I really think that for whatever reason... the scheduler is not as big a deal with the amd arch...
 
Patriot said:
huh... not the same experience I had... 1 smp was getting me ~13k on my 24core @ 1.7ghz

dual smp -12 were running about 10k a piece ...


I really think that for whatever reason... the scheduler is not as big a deal with the amd arch...
Click to expand...

That's very strange. I've never found on any architecture where running multiple SMP A3 (or later core) units made more points than running one at full usable threads.

Maybe with A2 a while back, but not since A3. Both my SR-2 and quad-box have been just fine going pedal-to-the-metal.

Only p670x units have shown that they just plain suck in terms of "return on electricity".

Regarding those units, I'd agree with what you'v said, but they are the exception.
 
New kernel to try:

http://en.fah-addict.net/news/news-0-349+new-bigadv-optimised-linux-kernel-available.php

Here is the second edition of our FAH-optimized kernel all you BigAdv addicts, based on version 2.6.37 of the Linux kernel. Frodo, our optimisation wizard, has hand-tailored this version specially for Intel Core machines.

List of features:
  • Special optimisations to take full advantage of Intel Core 2 / Core iX processor capabilities
  • Preemption mode enabled
  • 100Hz kernel timer
  • Scaling to support an unlimited number of CPUs
  • Performance mode (forces the i7's turbo mode to always be active)

Warning: Don't use this kernel if you're running a GPU client; this is meant purely for dedicated machines running only the CPU clients.

Please remember that this kernel is provided without any guarantees from the FAH-Addict team. It was tested on an Ubuntu 10.10 machine, and should work fine on Debian systems. We can't be held responsible for any issues caused, however unlikely. Install it only if you're capable of repairing your machine, or reinstalling your OS. Do not install on a production machine.

Your call is important to us.
Click to expand...
 
Never mind - I tried it, and same problem of not using all cores - frame time 11:42 - back to BFS with frame times of 9:54 please....

Oh well, worth a try....
 
I'm curious as to what these people are calling "optimized" kernels. Hopefully they left in /proc/config.gz support for me to find out. :p Looks like I have a weekend project now,
 
I've had a couple 6901s come in at 6:09 TPF, so 403K. Most come in around 6:15, so 394K. This is at 2.4ghz.

Properly configured, 2.5ghz should come in around 5:40-6:00 TPF (~400-450K).

tjmagneto said:
Let's see, with going from 2.3 Ghz to 2.5 Ghz it should be no problem.
Click to expand...
 
sfield said:
I've had a couple 6901s come in at 6:09 TPF, so 403K. Most come in around 6:15, so 394K. This is at 2.4ghz.

Properly configured, 2.5ghz should come in around 5:40-6:00 TPF (~400-450K).
Click to expand...

man i am depressed now ;)
 
Patriot said:
not renting a dell....sorry...
Click to expand...

So get your hands on an 8-processor ProLiant DL980 and run some tests for us! :p

I just customized one at HP.com and it came to $90K (software not included), or "Lease for as low as $2,352.61/month (48 months)". Not sure if I included enough RAM...
 
SazanEyes said:
So get your hands on an 8-processor ProLiant DL980 and run some tests for us! :p

I just customized one at HP.com and it came to $90K (software not included), or "Lease for as low as $2,352.61/month (48 months)". Not sure if I included enough RAM...
Click to expand...

I have a 2p 8core intel system ... but I am stinking out of ram...

It didn't like the network last time I tried to use it...
hmm... will screw around with it tomorrow... even if I have to borrow some ram from the lowest end folder...

BL 680
 
You must log in or register to reply here.
Back
Top