I just built a Q6600. I just installed the Folding SMP client on Vista.
I had some problems downloading the SMP client. So, in the meanitme, I installed the console client 4 times. This seemed to work fine.
I downloaded the SMP client at work and emailed it home. I turned off the 4 console services (set to manual start). The SMP client seemed to install fine. I went through the install, set my name and team, I selected No for service, yes for big WUs. Everything else default, including 100% utilization.
When I run FAH, I see the 4 cores bouncing between 40% and 60% utilization. Not exactly the 100% I expected. When I had the non-SMP clients installed, all 4 were pegged at 100%.
Is this normal, or did I mess something up?
SMP Question
Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee
-
peteamer
- *Lifetime Patron*
- Posts: 1740
- Joined: Sun Dec 21, 2003 11:24 am
- Location: 'Sunny' Cornwall U.K.
Hi David, it's normal, Linux SMP does the same thing.
IIRC it's something to do with cores waiting for all cores to finish their current bit before they all start together on the next bit/s.
Future clients are hoped to lessen this.
Pete
P.S. sorry the for poor explanation, England is my first language.....
IIRC it's something to do with cores waiting for all cores to finish their current bit before they all start together on the next bit/s.
Future clients are hoped to lessen this.
Pete
P.S. sorry the for poor explanation, England is my first language.....
-
NeilBlanchard
- Moderator
- Posts: 7681
- Joined: Mon Dec 09, 2002 7:11 pm
- Location: Maynard, MA, Eaarth
- Contact:
-
aristide1
- *Lifetime Patron*
- Posts: 4284
- Joined: Fri Apr 04, 2003 6:21 pm
- Location: Undisclosed but sober in US
David, are you seeing any checkpoints?DavidG wrote:I can't at this point. I think I've done something wrong. It looks like the cores are doing something, but it has not completed a WU since Mon night/Tue early morn (1 AM). It took about 18 hours to complete WUs this weekend with 4 instances.
I'm going to give it till tonight, then start over ...
Nope. I gave up and went back to 4 clients.aristide1 wrote:David, are you seeing any checkpoints?DavidG wrote:I can't at this point. I think I've done something wrong. It looks like the cores are doing something, but it has not completed a WU since Mon night/Tue early morn (1 AM). It took about 18 hours to complete WUs this weekend with 4 instances.
I'm going to give it till tonight, then start over ...
I'm going to wait unitl they release a version of the SMP client that includes a service install.
Each client is doing about 300-350 PPD.
I have heard of people running two instances of the SMP client on a quad core system, and setting the processor affinity to two cores for one and the other two for the other client.
I do not think that still gets them to 100%, but much closer.
I am getting more PPD than your 4 clients from my dual core E6700 running the SMP client, (regularly completing the 1760 pt work units in under a day) and you should be able to get close to twice as much as I am.
I do not think that still gets them to 100%, but much closer.
I am getting more PPD than your 4 clients from my dual core E6700 running the SMP client, (regularly completing the 1760 pt work units in under a day) and you should be able to get close to twice as much as I am.
So ... are you ocÃng?
I am curious, because the 1760 pt WUs take quite a lot longer (around 28 hours) on my machine, which is running a C2D E6750 at stock. I tried using my Gigabyte motherboard's CIA to boost the clock speed under load conditions; it only made the time-display run fast!
When a WU is running, the system activity display shows nice at about 65%, so I guess the CPU speed is not ramping up. Or the FAH program is more dependent on other component speeds such as memory and disk/io.
Regards
Mike
When a WU is running, the system activity display shows nice at about 65%, so I guess the CPU speed is not ramping up. Or the FAH program is more dependent on other component speeds such as memory and disk/io.
Regards
Mike
Hey Mike,
I have tried a few things that gained me little bits of time. Since I started out over 1 day per 1760, but close, I thought it would be nice to see a work unit appear each day. (I did not think I would be drawn into this, but the teams and ranked lists has motivated me to donate more points. Certainly good idea on their part.)
I also have a gigabyte board, and I am not at all a big overclocker, but I bumped mine up a little. I have mine currently running at 3030 with voltage bumped up to 1.35 and memory voltage bumped up to 2.0. (I see many people get much more overclock out of these, but I want to stay with basic heatsink and fan, and was not going to push the voltage to 1.4 as I have seen others do. I think I picked my voltage as half way between stock and what I saw overclockers talking about using.)
After doing this, I saw the time between steps go down in roughly the same percentage that I adjusted the clock.
At this point, I was just seconds per step over a day on the 1760's.
I had heard that the 64-bit SMP Linux client performed better than the Windows client. I still want windows on this machine for other reasons, but shaving a little more off the folding times would be nice.
A couple friends at work both thought that running the Linux SMP client under VMware might get me that. I was a bit skeptical that running a second OS on top of the other one could possibly be faster for folding, but I gave it a go.
I first installed free VMware server and normal Ubuntu 64 bit distro. That already put me under 1 day/1760, but was not all that much faster. Since I just wanted to fold, and did not really need Gnome or KDE, etc., I re-installed a 64-bit console only (or do they call it command line only) version, and have been folding at about 13 mins a step.
I have found that when I am not using the machine for anything else, I bump up the vmware task to above average priority (in Windows), and for some things, I bump it down to low so that it does not compete as much when I am using the machine. In that way, the VMware/Linux combination does seem to be a little more intrusive than the Windows client which I could just leave running and almost never noticed it was there.
Anyway, that summarizes all the things that I have tried or heard about to shave a little time per step off folding.
I have tried a few things that gained me little bits of time. Since I started out over 1 day per 1760, but close, I thought it would be nice to see a work unit appear each day. (I did not think I would be drawn into this, but the teams and ranked lists has motivated me to donate more points. Certainly good idea on their part.)
I also have a gigabyte board, and I am not at all a big overclocker, but I bumped mine up a little. I have mine currently running at 3030 with voltage bumped up to 1.35 and memory voltage bumped up to 2.0. (I see many people get much more overclock out of these, but I want to stay with basic heatsink and fan, and was not going to push the voltage to 1.4 as I have seen others do. I think I picked my voltage as half way between stock and what I saw overclockers talking about using.)
After doing this, I saw the time between steps go down in roughly the same percentage that I adjusted the clock.
At this point, I was just seconds per step over a day on the 1760's.
I had heard that the 64-bit SMP Linux client performed better than the Windows client. I still want windows on this machine for other reasons, but shaving a little more off the folding times would be nice.
A couple friends at work both thought that running the Linux SMP client under VMware might get me that. I was a bit skeptical that running a second OS on top of the other one could possibly be faster for folding, but I gave it a go.
I first installed free VMware server and normal Ubuntu 64 bit distro. That already put me under 1 day/1760, but was not all that much faster. Since I just wanted to fold, and did not really need Gnome or KDE, etc., I re-installed a 64-bit console only (or do they call it command line only) version, and have been folding at about 13 mins a step.
I have found that when I am not using the machine for anything else, I bump up the vmware task to above average priority (in Windows), and for some things, I bump it down to low so that it does not compete as much when I am using the machine. In that way, the VMware/Linux combination does seem to be a little more intrusive than the Windows client which I could just leave running and almost never noticed it was there.
Anyway, that summarizes all the things that I have tried or heard about to shave a little time per step off folding.
Thanks, I will give it a go
Thanks VanWAGuy
I guess it was lazy of me to think I could find a single control that would transform my GTI into a hotrod
.
Next time I have to reboot, I will give it a try. One major unknown is that I am running KDE the whole day as my window manager, and I don know how many cycles that soaks up. Nor can I guauge accurately the impact of running ktirrent at the moment, to spread the gospel of Mandriva 2008.0. But I guess thatÅ› no different to running VMWARE and Windows as well.
Time will tell.
Mike
I guess it was lazy of me to think I could find a single control that would transform my GTI into a hotrod
.Next time I have to reboot, I will give it a try. One major unknown is that I am running KDE the whole day as my window manager, and I don know how many cycles that soaks up. Nor can I guauge accurately the impact of running ktirrent at the moment, to spread the gospel of Mandriva 2008.0. But I guess thatÅ› no different to running VMWARE and Windows as well.
Time will tell.
Mike
keep us posted on stuff.
im anxiously awaiting a fix for HT processors. I have an old dual xeon box, and with hyperthreading turned on, it sees 4 cores, and so tries to act accordingly, but cannot finish a WU in time. So i just have it running several single threaded WUs
1 day for a 1700pt WU on a Q660? im jealous
I thought my dual opteron was doing nice at 3days per WU....
im anxiously awaiting a fix for HT processors. I have an old dual xeon box, and with hyperthreading turned on, it sees 4 cores, and so tries to act accordingly, but cannot finish a WU in time. So i just have it running several single threaded WUs
1 day for a 1700pt WU on a Q660? im jealous
I thought my dual opteron was doing nice at 3days per WU....
-
aristide1
- *Lifetime Patron*
- Posts: 4284
- Joined: Fri Apr 04, 2003 6:21 pm
- Location: Undisclosed but sober in US
Re: Thanks, I will give it a go
Wow, what a blast from the past that statement was. Had an 1984 GTI, did the exhaust, throttle body, cam. It was faster, but it was still no match for anything bigger.Dutchmm wrote:I guess it was lazy of me to think I could find a single control that would transform my GTI into a hotrod
Mike
There is no replacement for good old displacement.
-
NeilBlanchard
- Moderator
- Posts: 7681
- Joined: Mon Dec 09, 2002 7:11 pm
- Location: Maynard, MA, Eaarth
- Contact:
Hello,
I think the SMP client favors (in descending order):
Quad cores (probably Intel over AMD?)
Core 2 Duo
Athlon 64 X2
...and for OS's:
Linux/Mac OSX
Linux in VM Ware
Windows
The Core 2 Duo in Windows is faster than a X2 in Linux. I think an X2 (or dual Athlons?) in Windows is the slowest SMP setup?
A question: my Linux SMP is trying to download a new '78' core, and it cannot (overnight at least). It has been working like a champ for months, and I had recently updated to the Beta 6 SMP client, and it had worked fine for two work units. Has anyone else seen something like this?
I think the SMP client favors (in descending order):
Quad cores (probably Intel over AMD?)
Core 2 Duo
Athlon 64 X2
...and for OS's:
Linux/Mac OSX
Linux in VM Ware
Windows
The Core 2 Duo in Windows is faster than a X2 in Linux. I think an X2 (or dual Athlons?) in Windows is the slowest SMP setup?
A question: my Linux SMP is trying to download a new '78' core, and it cannot (overnight at least). It has been working like a champ for months, and I had recently updated to the Beta 6 SMP client, and it had worked fine for two work units. Has anyone else seen something like this?
Last edited by NeilBlanchard on Mon Oct 15, 2007 8:41 am, edited 1 time in total.
-
aristide1
- *Lifetime Patron*
- Posts: 4284
- Joined: Fri Apr 04, 2003 6:21 pm
- Location: Undisclosed but sober in US
Neil,
I don't think from a performance perspective that X2s can compete with Core 2 Duo's on any measurement. I don't think they were meant to. I think they hit the mark they wanted to hit, 70% of the performance at 30% of the cost, when I compare my X2 4000+ to my C2D E6400, $70 versus $210.
Typical NewEgg comments are "Its zips through what I need it to do and its lots cheap than Intel."
And of course the inevitable "IT ROCKS!"
There's a little area east of Syracuse called Solvay. It doesn't look like much (nothing in central NY looks like much) but they buy their electricity through some kind of coop program, and they typically pay half for electricity that everybody else pays. What a great place for a folding farm.
I don't think from a performance perspective that X2s can compete with Core 2 Duo's on any measurement. I don't think they were meant to. I think they hit the mark they wanted to hit, 70% of the performance at 30% of the cost, when I compare my X2 4000+ to my C2D E6400, $70 versus $210.
Typical NewEgg comments are "Its zips through what I need it to do and its lots cheap than Intel."
And of course the inevitable "IT ROCKS!"
There's a little area east of Syracuse called Solvay. It doesn't look like much (nothing in central NY looks like much) but they buy their electricity through some kind of coop program, and they typically pay half for electricity that everybody else pays. What a great place for a folding farm.
Neil,
Overall, I agree with your post, and nice brief summary by the way.
There might be an exception to the OS list though. I am not quite sure (and would like to have a quad to try) if Linux or Linux under VMware is best PPD for a quad core. I have seen a lot of messages that Linux on quad does not keep the CPU that close to 100% utilization. I have heard that some folks run 2 virtual machines, and the affinity for one is set to use CPU 0 and 1, and the other 2 and 3. I have heard conjecture that combination might keep threads scheduled on the same core more, resulting in less cache misses, or less traffic on the Front side bus.
Overall, I agree with your post, and nice brief summary by the way.
There might be an exception to the OS list though. I am not quite sure (and would like to have a quad to try) if Linux or Linux under VMware is best PPD for a quad core. I have seen a lot of messages that Linux on quad does not keep the CPU that close to 100% utilization. I have heard that some folks run 2 virtual machines, and the affinity for one is set to use CPU 0 and 1, and the other 2 and 3. I have heard conjecture that combination might keep threads scheduled on the same core more, resulting in less cache misses, or less traffic on the Front side bus.