New 2665 WU's

S

Sunin

[H]ard|DCer of the Month - August 2008
Joined
Dec 27, 2005
Messages
3,421
They are named HGG in water and wow are they the worst I have ever seen in my life!

I have a Quad 3.2Ghz getting a whopping 728.57 points per SMP it takes over 25 minutes per % and is only worth 1275. Pref and Final Deadlines are 5d +

Currently my 9 Quads are churning only 28k points. Ouch... I have 5 of these units churning and that represents 7500 points lost! Normally I see 2100+ points per with 2653's!

I'm really begining to question their benchmarks at F@H... there is no way a Pentium is churning this guy faster than my quads...



 
Want to rile up 7im? Go to the Folding Forum thread on this unit and ask him if this point reduction is Stanford's way of "telling" us they don't want us running SMP any more. :p



 
This isn't the first time and definitely won't be the last time points seem to be really messed up. Right now, considering the points bonus SMP work units get I wouldn't be surprised if they are messing with the point structure regarding SMP work units. The baseline machine for the quad work units is a Mac Pro or something like that.

If you want to hear something really bad, my previous system at work was a 2.8Ghz P4. It's the baseline system for the single core work units. Supposedly the baseline system with SSE optimizations turned off was supposed to get 110PPD or around there. With SSE optimizations turned on, I was lucky to hit 100PPD. That's messed up.

Right now, I have a Pentium [email protected] with two single clients running here at work. Regular gromacs work units might process at around 110-120PPD on this machine. I currently have two amber core work units crunching and they can't even push out 100PPD. Sadly enough, on my previous system I would get anywhere from 130-160PPD with amber core work units. I don't actually do much that is CPU intensive while working so frame times while idle and while the system is under work load are similar.

As a recap, it's sad that the benchmark system can't get near the stated PPD for the benchmark system. It's even sadder that a system better than the benchmark system can barely make the minimum PPD of the benchmark system at the best of times. This isn't anything new and I don't see any of this changing anytime soon.

 
SmokeRngs said:
This isn't the first time and definitely won't be the last time points seem to be really messed up. Right now, considering the points bonus SMP work units get I wouldn't be surprised if they are messing with the point structure regarding SMP work units. The baseline machine for the quad work units is a Mac Pro or something like that.

If you want to hear something really bad, my previous system at work was a 2.8Ghz P4. It's the baseline system for the single core work units. Supposedly the baseline system with SSE optimizations turned off was supposed to get 110PPD or around there. With SSE optimizations turned on, I was lucky to hit 100PPD. That's messed up.

Right now, I have a Pentium [email protected] with two single clients running here at work. Regular gromacs work units might process at around 110-120PPD on this machine. I currently have two amber core work units crunching and they can't even push out 100PPD. Sadly enough, on my previous system I would get anywhere from 130-160PPD with amber core work units. I don't actually do much that is CPU intensive while working so frame times while idle and while the system is under work load are similar.

As a recap, it's sad that the benchmark system can't get near the stated PPD for the benchmark system. It's even sadder that a system better than the benchmark system can barely make the minimum PPD of the benchmark system at the best of times. This isn't anything new and I don't see any of this changing anytime soon.

Click to expand...

Yes, someone need a slap on the head and a good full check on why they vary. Technically, if we have a boxen with similar specs of the benchmark machine, we should get the same PPD.

IIRc, the benchmark SMP machine is a Mac Pro with a clovertown cpu (maybe 2). They should know that running a Mac doesn't give the same results as a standard PC.

aldamon, go make fun of 7im ;)

 
Xilikon said:
Yes, someone need a slap on the head and a good full check on why they vary. Technically, if we have a boxen with similar specs of the benchmark machine, we should get the same PPD.

IIRc, the benchmark SMP machine is a Mac Pro with a clovertown cpu (maybe 2). They should know that running a Mac doesn't give the same results as a standard PC.

aldamon, go make fun of 7im ;)

Click to expand...


Yea, there doesn't seem to be enough anger over there yet. Here's the thread:

http://foldingforum.org/viewtopic.php?f=16&t=2561
 
I posted something and just stated that in the end you need to be consistent. This crap where we invest money only to have things drastically change, such as the 1950's and now it appears the Quads are heading that way... Cosidering the Bench machien is supposed to get 1760 PPD a Quad should average 2x that. I just don't get it... My Quad should not be getting 1500PPD that translates into about 8.5 PpWpD /slaps forehead wow... considering that costs me about .12 cents for that Point... I question this point system highly. They need to think about the rising costs to keep these machines running as well.



 
Sunin said:
I posted something and just stated that in the end you need to be consistent. This crap where we invest money only to have things drastically change, such as the 1950's and now it appears the Quads are heading that way... Cosidering the Bench machien is supposed to get 1760 PPD a Quad should average 2x that. I just don't get it... My Quad should not be getting 1500PPD that translates into about 8.5 PpWpD /slaps forehead wow... considering that costs me about .12 cents for that Point... I question this point system highly. They need to think about the rising costs to keep these machines running as well.



Click to expand...

I also added more infos. Since that WU have a 6 days deadline, the total points should be more than a 3065, which is benched at 2144 points.

While I didn't said on the FCF forum, I also think it's possible that this would be the first SMP WU without bonus due to the lax deadline ?

 
Maybe true with the lax deadline but then should we consider running Quad SMPs on a Quad? I mean at this level my GPU2 folding box is more attractive... which maybe it should be for the amount of science done. I still think they should pick a system and set all benches based on that system. I'd love to see a C2D try to churn the amount of science the PS3 can do! Just wish they would figure it out so we know where we can get the most bang for the buck if we are investing and keep it that way...

Consistency being key...

 
Sunin said:
Maybe true with the lax deadline but then should we consider running Quad SMPs on a Quad? I mean at this level my GPU2 folding box is more attractive... which maybe it should be for the amount of science done. I still think they should pick a system and set all benches based on that system. I'd love to see a C2D try to churn the amount of science the PS3 can do! Just wish they would figure it out so we know where we can get the most bang for the buck if we are investing and keep it that way...

Consistency being key...

Click to expand...

The reason given for the inconsistency is that the PS3 and GPU clients are indeed more powerful than the SMP, but are also less versatile in how many types of calculations that they can do. The SMP client can work many more types of problems, which is why it gets more point production.
 
I noticed something was wrong with one of my boxen on the way to work. One of the clients was only hitting 1000ppd on a [email protected] (WinXP dual SMP config with no VMs). I figured something hung for a 2653 to be screwed up like that. What do you know it is a 2665. 25hrs on a [email protected] for 1275 points seems very inconsistent.

 
I have not seen any of these yet ..

Do u have Advmethods on -- is that giving you 2665?
Mine of off so might be a factor since I havn't gotten one
 
This WU is a memory sucking pig. Better have at least 2GB if you want to do anything useful while its crunching. At least my Opteron 175 will make the preferred deadline, which it could not do with the 306* units. Frame times are around 34 minutes.
 
Well, I had a reply all typed up this afternoon at work when the forums crapped out. I didn't feel like retyping it so I emailed it to myself and here it is:

Sunin said:
Maybe true with the lax deadline

Click to expand...



Six days is not a lax deadline no matter how you look at it. A lax deadline is something along the lines of 60-90 days. Until the SMP client came out, a short deadline was something like 30 days. This was for a couple reasons. The projects were basically bigger and took longer to get all the results anyway and Stanford realized that this was supposed to be done with people's spare CPU cycles. To turn in a lot of the quad only units, the systems practically have to be folding and doing nothing else. I can barely make recommended deadlines on them while running dual Linux clients and I have that system overclocked by 50%. That's with doing virtually nothing else on the system.



The SMP client is still very much in beta; you just have to look at the Windows SMP client to see that. The Linux SMP client seems to be very stable for me, but it still has little problems and glitches occasionally. There is no reason not to have bonus points with the problems the client has. Also, as long as they plan on keeping work units with deadlines of less than 30 days there should be a bonus.



Myself and many others on this forum do a lot more than just let the clients run on systems using only spare CPU cycles. I have five or six machines running in my apartment right now. If it wasn't for DC projects, I would not have them. I build and run machines just for this. I think the least Stanford can do is actually make the damn points consistent. It can't be that fucking hard to do. They have benchmark machines and I want to know why in the seven levels of hell I can't get near the points the benchmark machine can do with a machine which is the damn benchmark machine and why a machine with faster hardware than the benchmark machine still can't match the performance of the benchmark machine.



These are many of the reasons I took all my single core machines off of F@H. They just aren't worth the time and power for me to run them on that project. The only exception has been my work machine because F@H is the only thing I can install on it. All my single core machines and most of my borgs run WCG instead since it seems they actually get something done on that project and receives work units suited to the speed of the machine.



 
I dunno but my production is down to 26kPPD because of these units... it is literally 1/3 the points, its sad when I wish for the 3062's and 3065's at least those I could avg about 3500 PPD per quad... I love how they are just avoiding the forums at F@H... No replies like we will look into it... etc..

 
What is the PPD on 2665 with all 4 cores crunching it instead of two intances? If Stanford is increasing the complexity of the units to snuff out dual instances, that's their choice I guess.

 
aldamon said:
What is the PPD on 2665 with all 4 cores crunching it instead of two intances? If Stanford is increasing the complexity of the units to snuff out dual instances, that's their choice I guess.

Click to expand...

After the Chimp Challenge I will convert a One Quad to a Single SMP and see what that does.
 
BillR said:
I've tried anger, common sense etc. They (stanford) have a listening problem.

http://foldingforum.org/viewtopic.php?f=16&t=2561&p=23477#p23477

Seems they dropped me from the rolls so I had to re-up;)

Click to expand...

and you were quickly brushed off by 7im

LOL I agree Bill you are dead accurate. Transparency, some explanation of setting expectations and such... I mean we do this for free... donating $1,000's of equipment to the cause and we'd like to be at least somewhat informed on what is going on. They should be proactive and say here is the issue, here is what you will see in the future, here is what you need to do.. .etc.. lay it out for us and they haven't. Hopefully they will. I'm not upset to the point of quitting the cause, but going from a high of about 37k to a low of 26k with the exact same hardware is kinda upsetting and knowing that my points if I was doing nothing but 2665's could bottom out at 13,500 PPD makes me feel a bit shafted considering the investment I have made.





 
Then I kicked 7im in the nuts :p

with things i know, I already replied to 7im's comment and basically asked Stanford to make a PUBLIC announcement explaining what they want to accomplish, the current issues and what they see from the future.

 
BillR said:
I've tried anger, common sense etc. They (stanford) have a listening problem.

http://foldingforum.org/viewtopic.php?f=16&t=2561&p=23477#p23477

Seems they dropped me from the rolls so I had to re-up;)

Click to expand...

You should get out more. You would know that there was a new folding forum. Everyone had to sign up again. ;)

@ X, Vijay doesn't usually wade in this deep, but since you asked so nice for a change, I will do what I can to get a comment from Stanford. I'd hardly call that civility a kick in the nuts.

And those of you on the folding beta team, you already know there is a silver lining on the other side of the p2665 cloud with the A2 fahcore. Wish I could say more. One of your own already hinted at this earlier in the thread. ;)
 
7im said:
You should get out more. You would know that there was a new folding forum. Everyone had to sign up again. ;)

@ X, Vijay doesn't usually wade in this deep, but since you asked so nice for a change, I will do what I can to get a comment from Stanford. I'd hardly call that civility a kick in the nuts.

And those of you on the folding beta team, you already know there is a silver lining on the other side of the p2665 cloud with the A2 fahcore. Wish I could say more. One of your own already hinted at this earlier in the thread. ;)
Click to expand...

I wish I could say more as well and that's why I really wanted a official post from the Pande group about that. We want to know where Stanford wish us to go and it's not a moderator, a user or someone outside who can make this kind of comments ;)

 
Xilikon said:
I wish I could say more as well and that's why I really wanted a official post from the Pande group about that. We want to know where Stanford wish us to go and it's not a moderator, a user or someone outside who can make this kind of comments ;)

Click to expand...

Let's wait and see what Stanford has to say. :cool:
 
7im said:
Let's wait and see what Stanford has to say. :cool:
Click to expand...
Hi 7im. I was wondering with all this talk about new and larger WUs for the SMP client, are we going to see 8-core support in the near future? I was thinking since the new P2665 WUs take long to process, wouldn't such WUs benefit from beefier hardware?
 
APOLLO said:
Hi 7im. I was wondering with all this talk about new and larger WUs for the SMP client, are we going to see 8-core support in the near future? I was thinking since the new P2665 WUs take long to process, wouldn't such WUs benefit from beefier hardware?
Click to expand...

Actually, it's a different issue ;)

 
Sunin said:
and you were quickly brushed off by 7im

LOL I agree Bill you are dead accurate. Transparency, some explanation of setting expectations and such... I mean we do this for free... donating $1,000's of equipment to the cause and we'd like to be at least somewhat informed on what is going on. They should be proactive and say here is the issue, here is what you will see in the future, here is what you need to do.. .etc.. lay it out for us and they haven't. Hopefully they will. I'm not upset to the point of quitting the cause, but going from a high of about 37k to a low of 26k with the exact same hardware is kinda upsetting and knowing that my points if I was doing nothing but 2665's could bottom out at 13,500 PPD makes me feel a bit shafted considering the investment I have made.






Click to expand...

I don't brush off quite that easy, I posted an answer;)



 
BillR said:
I don't brush off quite that easy, I posted an answer;)



Click to expand...

LOL i love the response it is dead accurate to my feelings on this. Just wish they would set the expectation once and leave it there. And expectations cannot be so losely worded that you have no idea what to expect. LOL that will make you go crossed if you think about that...

Wow I didn't realize how concerned this has made me... I was inches away from buying my 10th Quad and didn't. I want this issue resolved before I think of investing further, because maybe the thing to do is invest $100 per machine for a GPU client and get 2k pts that way versus another quad depending on what and how they perceive the science value... $400 gets me 4 GPU clients worth or 1 quad machine... 8k vs 1.5k? Wow.. um... going to see how this settles itself... I'm disappointed with F@H in the fact that I don't know what will provide the best bang for my buck anymore!

 
Sunin,

Bang for the buck is always going to be in a constant state of change. You are thinking of buying 4 ATI cards to get more points, but would you buy 4 NV cards instead if an NV client was just around the corner? Are you so loyal to NV that you would still buy NV cards even if the ATI cards performed better?s Or vice versa? For instance, I still see people buying AMD chips when the Intel chips perform better. What's up with that? Buying decisions aren't always just about points.

@ Bill, As to the original question, some of the transparency of client development is limited by it being in beta, but I agree that Stanford could do a better job with communications. Looks like Kasson from Stanford posted, but didn't answer your questions just yet. Hopefully what Xili and I have hinted at is that this problem will be short lived (actually turn in to a positive), and hopefully this will tide t[H]e team over until more information becomes available, either from kasson, or in the public forum in general.

@ Apollo - Yes, smp -n support is coming, up to 8 cores at the moment. We'll have to see what the 8 core 16 thread Nehalems with no Front Side Bus and tri-channel DDR3 memory brings us later this fall. This keeps the bang for the buck curve moving for Sunin too.
 
7im said:
Sunin,

Bang for the buck is always going to be in a constant state of change. You are thinking of buying 4 ATI cards to get more points, but would you buy 4 NV cards instead if an NV client was just around the corner? Are you so loyal to NV that you would still buy NV cards even if the ATI cards performed better?s Or vice versa? For instance, I still see people buying AMD chips when the Intel chips perform better. What's up with that? Buying decisions aren't always just about points.
Click to expand...

wow... technology changes? Really? Sorry for being sarcastic, but... Overall my point is I want to invest my money where it does F@H the most good, ie produces the most science which should correlate to the most points (NO?). Sorry I have 9 machines that fold 100% for F@H so yes my buying decisions are largerly based on the amount of science that I can accomplish for the cause for X amount of $$$.

Others it may vary because their buying a computer for say gaming and if the side benefit is they can fold great. For me on the other hand its purely a Science to $$$ thing. Well that and Science to Watt thing which translates into $$$ as Watts cost me $$$ :)



 
Sunin said:
wow... technology changes? Really? Sorry for being sarcastic, but... Overall my point is I want to invest my money where it does F@H the most good, ie produces the most science which should correlate to the most points (NO?).

Click to expand...

Yes, science should equal points, but not always or exactly equal. Stanford also uses the bonus to attract people to new beta clients. They might give a bonus if a new beta client is somewhat unstable to compensate you for a few lost work units. They may compensate you for doing specifically large work units, or work units that have large download sizes. You might be getting more points, but not necessarily doing more science, just helping in a different way that the project also values a lot.

Bang for the science buck, SMP is the way to go. Lower investment, lower ongoing costs, easier to maintain, easier to monitor, easier to upgrade, and produces a large, if not the largest amount of points, and does a lot of science very quickly. And run one fahcore per CPU core. Basically, the SMP client is the CPU client on steroids, and the CPU client is still the bread and butter of the project.

But don't get upset if that changes, again. ;) Unsubtle hint.
 
7im said:
Bang for the buck is always going to be in a constant state of change. You are thinking of buying 4 ATI cards to get more points, but would you buy 4 NV cards instead if an NV client was just around the corner? Are you so loyal to NV that you would still buy NV cards even if the ATI cards performed better?
Click to expand...

Actually, right now I would still buy nVidia video cards even if they were outperformed by ATI cards. Mainly because in my uses, ATI cards generally get their asses handed to them in Linux compared to nVidia. Until those driver issues with ATI change, no ATI for me.

The only way I wouldn't buy a new nVidia card is if they bring out another "dustbuster" FX5800. At that point I would just wait for the next generation.

Also, price can make a big difference in regards to building machines just for folding. My last three systems were Intel quads because I had the money and they had the best performance. If the current ATI folding client and work units had worked on the ATI cards back when I bought those systems, I may have changed my purchases. Back in November I purchased an X2 system due to some great deals I got. I wanted a quad to replace my E6400 at the time but I was able to build the X2 system cheaper than the price of a Q6600. I still get around 1350PPD out of the X2 running the SMP client. Basically, price vs folding performance can make a huge difference in some of our purchases.

As it sits, it's almost worth grabbing an ATI HD3870 to drop in my X2 system with the points they are putting out right now. It would require me to install XP on that system but I could run the GPU client as well as a single regular client for a nice increase in PPD over what the system is currently producing.

These are the reasons we need transparency and consistency. Some of us are forking out thousands upon thousands of dollars to build folding rigs. Some others aren't spending as much since they can't afford it but they want to get the most out of what they can afford.

 
to add some fuel to the fire, here are some numbers from a q6600 at 3.4GHz, Vista64, 1 instance of the 5.91 beta client:

project 2653's were running at ~8.5 minutes per 1%, FahMon PpD was ~2980

project 2665's are running at ~12.5 minutes per 1%, FahMon PpD is ~1470

the 2665's are using 94-140 MiB's of RAM, while the 53's were sitting lower at ~65MiB's IIRC


so for approximately 1.5x the time you get half the points

and for what it's worth, Stanford sure as hell needs to respect the people devoting time/energy/computers to do their work. without the people in the f@h community selling them their computing time for only internet points in return, they'd be stuck on their asses trying to get funding for a supercomputer. So that said, 1 billion internet points to everyone that sends me a new/slightly used Power6 rack :p
 
novadude said:
... without the people in the f@h community selling them their computing time for only internet points in return, they'd be stuck on their asses trying to get funding for a supercomputer. ...
Click to expand...

You're preaching to the choir around here. :D


 
7im said:
Buying decisions aren't always just about points.
Click to expand...

I have 2k in my basement that says differently.

Many of us have made purchases so we can get the most points. We do this because we think (logic is funny, hu?) what gives us the most points is what they want done for the project.

If they change the SMP clients around like this and throw in a lemon WU (compared to the others) what good does this do but start threads like this. :confused:

Can they not keep the masses (us) happy and give out points based on the science we do and the value of the WU?

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