Folding Forum

Did anyone notice that over the last few weeks/months the total number of active folders (as shown on http://folding.stanford.edu/home/blog#) has been dropping by ~1-2K / day? End of last year we were some 250K, today it has dropped below 190K.
That doesn't quite match with the idea to 'reach 1,000,000'.

I did notice it but what I would like so see is how the Petaflops are being effected; is it decreasing, increasing, constant. You can read some of the discussion about the decline in active folders here (viewtopic.php?p=256480#p256480).

VijayPande wrote:

7im wrote:...Could be people don't like V7. Could be the crappy economy. Could be everyone is moving to mobile devices. Could be an alien invasion. Could be new GPU fahcores. Could be an AMD stats glitch. Could be the increasing cost of electricity. Could be Global warming, so they turned off their FAH heaters...

Regarding the client count: we have had a large company donating computer time anonymously and that donation time naturally ran its course (they don't care about BA, etc). That covers about ~30,000 CPUs or so. It's unfortunate timing that that ended around the new year, coincidentally with some of the rough server backend issues we had and the BA discussion.

I am hoping that this group will let us publicly acknowledge their contribution soon as what they've done (and the work we've been able to do on those machines) has been pretty exciting for us.

A full explanation is impossible to determine, although there are a lot of pet theories floated here.

PantherX wrote:I did notice it but what I would like so see is how the Petaflops are being effected; is it decreasing, increasing, constant.

Within the same type, FLOPS = constant * #active_clients, meaning if where's a 50% drop in example Windows-computers, where's also a 50% drop in Windows-FLOPS.

This means, that appart from the adjustment done 26.02.2013, the reported FLOPS doesn't tell anything about users switching from older computers to faster computers, or slower GPU's to faster GPU's, since as far as Folding@Home's stats-page is concerned, it's no difference between an old Ati-HD5450 and a new Amd-9R-290X, even the latter probably is doing 50x - 100x more work/day.

Doing a little calculation, to produce one TFLOPS you need:
370 Mac-clients*
320 Linux-clients*
229.8 windows-clients
7.0 Amd-clients
5.3 Nvidia-clients
2.95 Fermi-clients

It's interesting how a single-core Windows-computer has no problem outcrunching a 16+ core BigAdv-Linux-computer...

*: Due to TFLOPS not including any decimals and Mac and Linux having so few computers, there's large variations in the calculated #clients/TFLOPS for these two platforms from day to day, but would expect it's within +- 20 clients/TFLOPS.

I would say that the number going down, is in response to the BA debate. I have a AMD 4p - 24 core @2.8 Socket F Server. I spent over $1200 to be able to fold BA and get 275K - 305K for ~20+ hours of Work (I was diagnosed with Leukemia in Aug 2012). I was going to add a 2nd server AMD 2p - 32 core @2.4 G34 server BUT not now. When my 4p does a SMP unit and I end up with 45K ~ 70K for 20 hours of work. It's cheaper to leave it turned off as the PPD to Watts RATIO SUCKS. Think About it. Companies that buy Servers/Clusters look at TOTAL WORK DONE = HOW MUCH ENERGY IT COST TO DO THAT WORK!!!!!!

SCREAMING

We have to wait 5 hours, Were not going to last 5 seconds. If you didn't notice, We just got our asses kicked.
Game over.....MAN
Game.......over

From the Movie Alien

Thank You For Your Time

MaddMutt

I've shut down four old dual-core machines and replace them with one GPU. Higher PPD and fewer things to track, so the number of machines I'm running is going down.

I think many others are doing what codysluder is doing as well. The FLOPs is looking very high even though the # of CPUs is going down.

Personally I have reduced the number of clients I have running, but increased my output.. shut down 2 old or low powered laptop/netbook/uniprocessor and running gpu and 6 core smp. When I started the project many years ago I built a 4x gpu (gt 9800) and 4core Phenom 920 rig, this was really pumping at almost 20kppd putting me in the 'most powerful folders' category. Now my GTX 780ti does more than 10 times that by itself.

I'm wondering if the metric of the total # of CPUs is a poor one to put on our main page. It's the wrong incentive -- we want the flops up not the total # of CPUs. When donors like kiore and codysluder upgrade FLOPs (even if their CPU count goes down) that should be seen as a positive change for FAH, but now it's seen as a negative (if we highlight CPU count).

VijayPande wrote:I'm wondering if the metric of the total # of CPUs is a poor one to put on our main page. It's the wrong incentive -- we want the flops up not the total # of CPUs. When donors like kiore and codysluder upgrade FLOPs (even if their CPU count goes down) that should be seen as a positive change for FAH, but now it's seen as a negative (if we highlight CPU count).

As long as FLOPS = constant * CPU-count as is the case at the moment, the only way to increase FLOPS is by increasing #CPU's (or #GPU's).

It also means where's currently no difference between a single-core CPU and a 32-core CPU as far as the FAH-FLOPS-count is concerned, as long as both CPU's is running the same OS.

As for only highligting FLOPS on the front-page, this is of course an option, but I'm not sure if it's the best metric if FAH isn't planning to become a NVIDIA-only project.

That could be true as far as total computational power goes but is not the goal to have more active members doing said work, does flops accurately measure active donors. It would seem to me it does not perhaps you could provide both active members and total flops. Seems to me more active members = more flops.

What he said.

FLOPS increases over time, simply has a function of hardware upgrades. Even if the client count was constant, the FLOPs would keep going up.

With every client being a multi-cored system these days, client count is a good start, but active cores needs to be included, even if half that number is hyperthreads. It also doesn't account for the number of cores that GPUs bring to bear on the topic. How do you represent the difference between a low end GPU with 200 shaders, and the high end with 2000 shaders?

As always, no simple answer covers it.

I believe that total # of CPUs still conveys relevant information. On the surface, # CPUs also allows us to make some approximations such as the average computing power of each OS/hardware class. Its interesting to know information on donor statistics and computing power.

Using the example of codysluder, turning off 4 dual core machines in favor of one GPU may increase the total number of FLOPs but it is a smaller donation of resources in terms of electricity, effort and total hardware. While this may or may not be occurring on a larger scale I think that its hard to draw any meaningful conclusions based on one piece of anecdotal evidence. Does PG have any additional data on users/clients clients that can be analyzed to tease some more concrete evidence regarding CPUs vs FLOPs?

In reality, # CPUs is only an issue at the moment because it is going down. I don't think providing less information is the answer, especially in the current folding environment. How the information is presented can have a large impact on how users perceive it. Perhaps change the display to emphasize FLOPs rather than eliminating the number of CPUs would achieve the intended result without losing additional information.

Thanks. We wouldn't remove the CPU count on the osstats page. Just switch to FLOPs on the main folding.stanford.edu. The osstats page would be unchanged and no info would be lost. However, I think the current setup is giving a misimpression in terms of how FAH is doing.

In the case of displaying a single number, FLOPs most likely provides a better summary that # CPUs.