Folding Forum

The official stats keep track of two numbers: The number of points earned and the number of WUs completed. Many years ago, the variation in WU complexity was much narrower and WU-count was a good measure, but as the complexity range widened, points became more important, followed by the addition of a measurement of speed which was incorporated into QRB. Although almost everybody who compares their results to someone else's uses points, but there's no good reason to stop counting WUs

Comparing that to today's discussion, I think we need to track both the total number of clients and the total number of flops (with a reference to the discussion of one kind of flops vs. another kind).

Adding high-performance clients is good, and it'll show up on the flops count. Adding low-performance clients is also good, and it'll show up on the client-count number.

Rattledagger wrote: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).

but that's not the case. When people switch between machines and/or to/from GPUs that makes a difference in FLOPs/machine and so the total CPU count is misleading for that reason.

When did anybody ever upgrade their i7 to a P4 (or equiv.)?

ok, I think it makes sense to find a way to show both numbers (flops & raw count) more prominently.

VijayPande wrote:but that's not the case. When people switch between machines and/or to/from GPUs that makes a difference in FLOPs/machine and so the total CPU count is misleading for that reason.

Just to be absolutely clear, the FLOPs reported on the OS Stats page is an actual measurement of the FLOPs of each client? So upgrading from two Core i3's that each were contributing 10 GFLOPs to a Core i7 that contributes 100 GFLOPs (both figures made up for simplicity) would result in a net gain of 80 GFLOPs as measured on the OS stats page?

Also, while I don't know how easy it would be to implement, I think it would be useful to have an official number of active donors reported. There's been some speculation about that recently, and the best people could come up with were counts from Kakao (34k active users) and EOC (14k active users, with the caveat that EOC doesn't process everyone). That's somewhere where a lull in corporate donations may not look as drastic. I don't know how those corporate donations were counted, but if they were all folding as 'anonymous', then when they ended the active client count may go down a lot but the active user count would not. It seems a little odd to not have that figure readily available.

Zagen30 wrote:Just to be absolutely clear, the FLOPs reported on the OS Stats page is an actual measurement of the FLOPs of each client? So upgrading from two Core i3's that each were contributing 10 GFLOPs to a Core i7 that contributes 100 GFLOPs (both figures made up for simplicity) would result in a net gain of 80 GFLOPs as measured on the OS stats page?

Yes.

http://fah-web.stanford.edu/cgi-bin/mai ... e=osstats2

The FLOPS per core was last updated based on a FAH core performance report run on Tue, 26 Feb 2013...

*TFLOPS is the actual teraflops from the software cores, not the peak values from CPU/GPU specs. Please see our FLOPS FAQ...

If you want to know if there are more or less people folding.
Would it not be better to count the names of actual people folding, and not core count or FLOPS?
You can have less people folding and have higher core count or more FLOPS depending on what systems they bring online.
What number of actual people were folding January 29 2013 vs. January 29 2014?

Bill1024 wrote:If you want to know if there are more or less people folding.
Would it not be better to count the names of actual people folding, and not core count or FLOPS?
You can have less people folding and have higher core count or more FLOPS depending on what systems they bring online.
What number of actual people were folding January 29 2013 vs. January 29 2014?

Names are not unique. Active Passkeys?

And what counts as Active? Over what time period? If a P4 can still complete work units by the 20th day deadline, then shouldn't we count that as active? What is the longest WU deadline at the moment? (59 days Timeout, 88 days Deadline) Do we use the Timeout or Deadline?

But is someone who turns in one WU each month really "active"? Maybe we need two time periods? 2 weeks and 2 months?

The stats system already tracks active clients per user in the last 7 days and the last 50 days. Why not take a count of total users who have at least one active client in 7 days and 50 days?

I realize that doesn't address multiple people using the same name, but I think that would be at least a ballpark estimate. If there's a way to track active clients per passkey that would probably be more accurate.

VijayPande wrote:

Rattledagger wrote: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).

but that's not the case. When people switch between machines and/or to/from GPUs that makes a difference in FLOPs/machine and so the total CPU count is misleading for that reason.

Short answer, read the stats-page and you'll see the text: "The FLOPS per core was last updated based on a FAH core performance report run on Tue, 26 Feb 2013."

BTW, I seems to be coming to the opposite conclusion based on this statement to 7im interpretation of this.


Longer answer:
If you takes a look on the historic stats, as saved on the google-sheet https://docs.google.com/spreadsheet/ccc ... ring#gid=0

you can very easily calculate for example Windows active_cpu/TFLOPS (for windows) and you'll get, from 26.02.2013 to 31.12.2013 the average was 229.8 active_cpu/TFLOPS, with a daily variation of 0.2 active_cpu/TFLOPS due to roundoff-errors since TFLOPS doesn't have any decimals.

If you do the same for Linux or Mac, where's so few Linux/Mac-computers that you'll have much larger daily variations, but you'll get a different constant. Similar for the various GPU's, for Fermi-NVIDIA you'll get 2.95 active/TFLOPS. For clients/TFLOPS for all platforms, see my 18.01.2014-post earlier in this thread.

As a comparison, for windows for the same period core/cpu has changed from a low of 1.40 core/cpu upto a top of 2.03 core/cpu but still it's at 229.8 active_cpu/TFLOPS +- 0.2 active_cpu/TFLOPS. Also, the minimum was 123707 active_windows_cpu and the max was 315317 active_windows_cpu. With so large variations in core/cpu I find it unrealistic this doesn't have any noticeable effect in active_cpu/TFLOPS if where's not a linear conversion-factor between active_cpu and TFLOPS.


re-edit - just to be clear, switching between cpu and GPU does give an effect, but switching between an old single-core windows-computer to a 32+ core windows-computer has zero effects on the TFLOPS as reported on FAH's "Client statistics by OS". (*)

(*): atleast according to my calculations. Anyone finding any errors in the calculations?

Rattledagger wrote:snip...

edit - just to be clear, ... but switching between an old single-core windows-computer to a 32+ core windows-computer has zero effects on TFLOPS.

How can you say that switching from old Pentium 4 does the exact same amount of FLOPS per core as a new i7 Quad core?

The speed of the SSE in the i7 is at least double the P4, so the FLOPS/core is at least double.

7im wrote:

Rattledagger wrote:snip...

edit - just to be clear, ... but switching between an old single-core windows-computer to a 32+ core windows-computer has zero effects on TFLOPS.

How can you say that switching from old Pentium 4 does the exact same amount of FLOPS per core as a new i7 Quad core?

The speed of the SSE in the i7 is at least double the P4, so the FLOPS/core is at least double.

Ah, did I miss the "based on the statistics as presented on FAH's 'Client statistics by OS'-page"

Or, do you disagree with my calculations of active_cpu/TFLOPS?

I'm fully aware of an i7 being significantly faster than an old p4, but atleast my calculations indicates this difference is not accounted for on FAH's statistics-page's TFLOPS-score.

Calcs are correct. The numbers you are using, and what you think they say is not.

The FLOPS per core is a derived number, based on a fixed value, set almost a year ago.

The FLOPS per core was last updated based on a FAH core performance report run on Tue, 26 Feb 2013.

So if all you are doing is taking the total FLOPS, and dividing by the total cores, you are simply deriving the exact same number, hence it never changes.


But this demonstrates the Stats page is not accurate. For one, it doesn't account for the Core 17 FLOPs speed with the date of that FLOPs report.

If we flip around RD's calculation and do GFLOPs/active CPU, you get around 4.35 GFLOPs going back over a year (it oscillates between around 4.347 and 4.355). I looked up some theoretical numbers, and even a Pentium 4 was apparently theoretically capable of 10 GFLOPs, while a modern entry-level dual core can do around 40 GFLOPs or so. I realize theoretical FLOPs is going to be higher than actual FLOPs, and not everyone folds 24/7, but does that low an average seem right? I would think that several years of hardware upgrades and new computers would raise the average up to more than that.

I looked at the FLOP's numbers: http://fah-web.stanford.edu/cgi-bin/mai ... e=osstats2 and came to some startling conclusions.

1. 96% of the FLOPS are being done by GPU's and only 4% by all CPU's.
2. only .08% of FLOPS are done by Linux and that includes all bigadv machines.

My conclusions:

If the number of clients dropping is not a concern because the FLOPS are staying consistent then obviously people are dropping large quantities of CPU slots in favor of a few GPU slots in droves.

If the measure of scientific work done is the FLOP then CPU's should have little value and bigadv is equivalent to none. The bigadv program being dropped is totally insignificant to the F@H except at the level of PR.

PPD's do not accurately the measure of scientific work/value.