Folding Forum

I was starting from the back! Not a big deal-- I'll go back to the front.

I'm doing it that way (half and half).

I actually liked it with the dots; it made it apparent which time frames had more/less data than others..

Thanks for helping!

Great stuff, I'll leave it to the two of you at this stage to stop shoe stepping, but what a great find!

I've been searching high and low for data around the end of 2011/start of 2012 as it will be beneficial for finding the 'peak' NH Winter folding rate.

k1wi wrote:Great stuff, I'll leave it to the two of you at this stage to stop shoe stepping, but what a great find!

I've been searching high and low for data around the end of 2011/start of 2012 as it will be beneficial for finding the 'peak' NH Winter folding rate.

Maybe in here: viewtopic.php?f=16&t=21317&start=30#p216008

Are we keeping track of where we're getting the numbers from? The Active ATI GPU count for 12/12/2010 is way off what it should be, and if I knew where the number came from we could just check again and fix it. In general, it might be useful if we add comments to the sheet as to where we got that line of information. Or maybe it's just me starting to fundamentally want citations attached to stuff...

The issue with the Active ATI GPU count for 12/12/2010 is from an accidental copy and paste of the Linux active CPUs.. Will see if I can find a corrected number.

Edit: Found and fixed it.

There are multiple records on 5/24/2009 but they are all different. Hmm.

I also added some dates and links to sources but didn't put in the data... yet.

Official reported TFLOPS dropped below 6 x86 PFLOPS

One thing I find interesting is that while Windows TFLOPS have fluctuated over time, since 2008 the 'peaks' have only increased by 50% over the 4 years, from 220TFLOPS to 330TFLOPS. Over that time windows participation rates has almost doubled, from 260,000 to nearly 500,000 active clients.

I believe that figure measures real FLOPS and not theoretical FLOPS, so accounts for both improvements in software 'efficiency' and hardware improvements.

Interestingly Linux clients appear remain on longer into the 'Northern Hemisphere summer' but then decreased quite rapidly. Windows units have been declining since November? I almost wonder whether some of the decline is due to the V7 client...?

Anyone know why NVidia GPU stats don't want to play nice with the per client graph?

k1wi wrote:Anyone know why NVidia GPU stats don't want to play nice with the per client graph?

I'm not entirely sure what you mean, but if its the numbers above the graph that I saw last month, that's Google Docs trying to add annotations. Idk why it tries to do that though...

I mean I haven't been able to get the NVidia GPU x86 stats into the page - adding the additional column doesn't result in the line being 'accepted' - no draw and no apparent options for the line. I don't have enough experience with google docs to make the appropriate mods to it but if someone else does then it would be great to see it included.

Does anyone know / have a way to message the guy who made this:

http://en.wikipedia.org/wiki/File:FAH-tflops.PNG

It covers a year that's pretty sparse on our chart, but he doesn't have a Wikipedia User page so I can't leave him a message. Thoughts?


Edit: I haven't touched the Per Client graph so I don't know what to say about that. A lot of it looks odd to me at the moment but I don't know enough about Google Docs to do anything about it.

His last Wikipedia contribution was in 2010. You can always create his Talk page and send him a message that way. Worth a shot, though it may not work. At worst case scenerio, you can always take the image, paste it into MS-Paint or the like, and look at the pixel coordinates and figure out the values from there. But that would be painful...

screen317 wrote:Does anyone know / have a way to message the guy...

Yes. I am a member of the forum where kaleb_zero used to reside. That donor stopped posting in that forum about the same time as in WIKI. No Good-bye posts, just stopped at the end of 2010.

I sent a PM asking if the data was still available. Hopefully we get a response. (and there is hope, as someone contributed some fah WUs to that account in early 2012.)

Would be great if we were able to get some data from that time period.

I'm wondering how FLOPs is calculated, in particular, do multi-day WUs reduce the daily FLOPs count? That is, does a folder folding once every two days report 0 FLOPS on the days they don't return WUs? If so then the length of WUs could severely influence the shape of the graph. Is FLOPs calculated in some other way, such as taken as the average output multiplied by the number of active folders?

The other point I have pondered at is the different measurements of 'Active' participation - for CPU clients the time period is 50 days, whereas it is 15 days for PS3s and only 7 days for GPUs (reflecting the smaller WU deadlines), one could suggest that uni-proc deadlines are much different to SMP deadlines now (although projects for the two are being unified). The difference in time periods will, in my estimation, make the participation rate fluctuate differently amongst different hardware (the graph will respond to GPU changes much more rapidly). By measuring 50 day active rates the participation graph will effectively miss all the people shutting down for a month over the peak summer heat.

I seem to be asking more and more questions!

As a side note, as of two weeks ago all of the HPCS private beta clients (that were shut down in time for the shift to paid service) are no longer measured as active.

WUs have an internal tally of the FLOPS count and is reported with each WU. So yes, and no. Multi-day WUs would reduce the the count, if only a few WUs were running. But with thousands running, it averages out so no drop. Kind of like your PPD averages out over time.

The Active Client length, as you said, will affect the graph, but only tends to shift the result. If the project stopped tomorrow, the graph for the GPUs would drop to zero in 7 days, while the PC graph would stay at the current level for 50 days before dropping off. But the reverse is not true. If the GPU client count doubled tomorrow, the graph would go up tomorrow.

It also hides any outages that only last a few days, or weeks, as you noted.

Overall, it only really smooths the curves in the long run, but not much of a big affect, because we don't often have big changes in participation.

Would be interesting to look at the chart from when the GPU1 client was shut down...

Its sort of like asking if multi-day WUs change your PPD. When averaged over a period of time, it doesn't matter if you completed 1 long WU or many short ones since the production rate of your hardware doesn't actually change much except if your client is paused. Technically, FLOPS is an instantaneous measurement which the server can't detect, but it does collect the data containing the total amounts of work completed whenever results are uploaded. Add them up for a month and it's pretty nearly constant.