4P reality: peak vs. avg performance

A

AndyE

Limp Gawd
Joined
May 13, 2013
Messages
276
Outing myself, I have to confess that hanging around here at the [H], the 4P bug bite me a while ago. :)

Obviously, I was intrigued by an "1 mio ppd i4P" system and so it happened to be built based on all the excellent info available here.

Due to the heat over here, the only system running in August was this one and I'd like to share the "real" average performance my system produced over a 14 day period vs. the 1mio ppd "target".

Within this 14 day period the system produced (produced, as delivered and recognized by the FAH servers) 28 WUs and 11,579,993 points, averaging over this period 827k ppd. This is about 17,3% less than the hoped-for (peak) performance. Checking the log files, the system got the following assignment distribution (28 WUs in total):
P8101 = 9 (32%)
P8103 = 7 (25%)
P8104 = 5 (18%)
P8105 = 7 (25%)

It would be great, if other 4P users could share their peak vs avg numbers as well, so I get a better understanding if the 17.3% "gap" in my case is within the usual range.

Thanks,
Andy
 
AndyE said:
Outing myself, I have to confess that hanging around here at the [H], the 4P bug bite me a while ago. :)

Obviously, I was intrigued by an "1 mio ppd i4P" system and so it happened to be built based on all the excellent info available here.

Due to the heat over here, the only system running in August was this one and I'd like to share the "real" average performance my system produced over a 14 day period vs. the 1mio ppd "target".

Within this 14 day period the system produced (produced, as delivered and recognized by the FAH servers) 28 WUs and 11,579,993 points, averaging over this period 827k ppd. This is about 17,3% less than the hoped-for (peak) performance. Checking the log files, the system got the following assignment distribution (28 WUs in total):
P8101 = 9 (32%)
P8103 = 7 (25%)
P8104 = 5 (18%)
P8105 = 7 (25%)

It would be great, if other 4P users could share their peak vs avg numbers as well, so I get a better understanding if the 17.3% "gap" in my case is within the usual range.

Thanks,
Andy
Click to expand...


What exactly are you running?
 
AndyE said:
Outing myself, I have to confess that hanging around here at the [H], the 4P bug bite me a while ago. :)

Obviously, I was intrigued by an "1 mio ppd i4P" system and so it happened to be built based on all the excellent info available here.

Due to the heat over here, the only system running in August was this one and I'd like to share the "real" average performance my system produced over a 14 day period vs. the 1mio ppd "target".

Within this 14 day period the system produced (produced, as delivered and recognized by the FAH servers) 28 WUs and 11,579,993 points, averaging over this period 827k ppd. This is about 17,3% less than the hoped-for (peak) performance. Checking the log files, the system got the following assignment distribution (28 WUs in total):
P8101 = 9 (32%)
P8103 = 7 (25%)
P8104 = 5 (18%)
P8105 = 7 (25%)

It would be great, if other 4P users could share their peak vs avg numbers as well, so I get a better understanding if the 17.3% "gap" in my case is within the usual range.

Thanks,
Andy
Click to expand...


What exactly are you running?

I also noticed you pulled more 8101s than any other WU which would obviously affect the data in a negative way. I'd be interested in seeing the data in another 2 weeks as I've gone weeks before without a '01 and have known some to go for days straight with ONLY 8101..

Sadly my average will never meet my expectations lol...
 
m33pm33p said:
What exactly are you running?
I also noticed you pulled more 8101s than any other WU which would obviously affect the data in a negative way. I'd be interested in seeing the data in another 2 weeks as I've gone weeks before without a '01 and have known some to go for days straight with ONLY 8101..
Click to expand...

I am running a 4x E5-4650 system with the classic [H] install script.
And btw: my system consumed 615 watt average over the 14 days (wall outlet).

Of course the # of P8101 have a huge impact on average performance - which was my point I tried to make. One one side we talk about a "1m ppd system" but due to the WU distribution, it will "never" be a real 1m ppd system :)
The "real" work produced by the system is lower.

So, this is my question:
Given all the different projects people are grabbing up with their 4P systems, what are the kind of "real" = average ppds you get over a longer period of time? And if the data is available, what is the percentage of 8101 WUs you are grabbing "voluntarily" :)

Sadly my average will never meet my expectations lol..
Click to expand...
.... so true ... :D
 
AndyE said:
I am running a 4x E5-4650 system with the classic [H] install script.

Of course the # of P8101 have a huge impact on average performance - which was my point I tried to make. One one side we talk about a "1m ppd system" but due to the WU distribution, it will "never" be a real 1m ppd system :)
The "real" work produced by the system is lower.

So, this is my question:
Given all the different projects people are grabbing up with their 4P systems, what are the kind of "real" = average ppds you get over a longer period of time? And if the data is available, what is the percentage of 8101 WUs you are grabbing "voluntarily" :)
Click to expand...

Man... It's a lot more fun to say I have a 1m ppd system though.... Who wants to say "weeeellll on average it's ONLY xxxx" ... :D
 
m33pm33p said:
Man... It's a lot more fun to say I have a 1m ppd system though.... Who wants to say "weeeellll on average it's ONLY xxxx" ... :D
Click to expand...
Looks like you are a marketeer :D
(I'd see myself more in the engineering department) ... :)
 
AndyE said:
Within this 14 day period the system produced (produced, as delivered and recognized by the FAH servers) 28 WUs and 11,579,993 points, averaging over this period 827k ppd. This is about 17,3% less than the hoped-for (peak) performance. Checking the log files, the system got the following assignment distribution (28 WUs in total):
P8101 = 9 (32%)
P8103 = 7 (25%)
P8104 = 5 (18%)
P8105 = 7 (25%)

It would be great, if other 4P users could share their peak vs avg numbers as well, so I get a better understanding if the 17.3% "gap" in my case is within the usual range.
Click to expand...

AndyE, I have 33 WUs I can draw from on my 4xE5-4650 system. My average PPD appears to be 928826.5. I had the following WU distribution:
P8101 = 10 (30%)
P8103 = 11 (37%)
P8104 = 5 (15%)
P8105 = 7 (21%)

Did you perform these calculations manually or did you figure out how to extract the data from HFM? I was using Work Unit History and for some reason I had WUs show up multiple times for the same host. I also couldn't quickly determine if HFM allows exports to CSV or how I could get the data into Excel for further analysis, but I didn't spend much time looking.
 
brilong,
thanks for sharing your data.

With your WU distribution, the avg ppd should be rather in the 870k range (The 30% P8101 pull the avg result down). The P8101 is between 570 and 610k ppd, the others are between 930 and 1010k ppd. With 30% P8101 share, an average of 929k ppd is beyond my system's capability.

I don't overclock, the CPU's are thermally stable (= they don't throttle) and I use 1600 MHz ECC RAM.

I used the "ultimate" metric - the final recognized points and did not rely on any locally computed values.

Here is a screenshot of the 14 day period of this system
original.jpg

To avoid any influence I extracted manually all data from the logfiles.

Typical times:
P8101 = 9m20s plus 2min transfer time for the 90mb datafile
P8103 = 6m57s plus 2m transfer
P8104 = 5m02s plus 2m transfer
P8105 = 6m54s plus 2m transfer

Andy
 
AndyE said:
brilong,
thanks for sharing your data.

With your WU distribution, the avg ppd should be rather in the 870k range (The 30% P8101 pull the avg result down). The P8101 is between 570 and 610k ppd, the others are between 930 and 1010k ppd. With 30% P8101 share, an average of 929k ppd is beyond my system's capability.
Click to expand...

I might have miscalculated. To be honest, I made a quick run through HFM.net's expected PPD per WU, so this calc. was not as good as yours.

AndyE said:
I don't overclock, the CPU's are thermally stable (= they don't throttle) and I use 1600 MHz ECC RAM.
Click to expand...
I overclock (all-cores 3176MHz according to i7z). 100% c0 state with temps up to 73C on hottest CPU. I'm using 1333 MHz ECC RAM.

AndyE said:
I used the "ultimate" metric - the final recognized points and did not rely on any locally computed values.

To avoid any influence I extracted manually all data from the logfiles.
Click to expand...

Did you write a script to parse all your logs or did you just trudge through it manually? I'd like to run the same calc against my logfiles if possible so we're comparing apples to apples.
 
Using EOC stats to calculate ppd is not going to be accurate either. Unless you start folding at midnight on the first day of the 2 week period and turn in your last WU at 11:59 PM on the 14th day, your total points would need to be averaged over something less that the 14 days you used. I think that the best way to calculate actual average ppd is going to be using HFM's reporting tool. I'll see if I can put something together once I get home tonight.
 
musky,
this is what I did.

As I have currently only one rig running, it was easy to start it on the hour of this period.
The last WU I considered came in 2 hrs past the 2 week period (with 354,587 points). As this unit took about 12 hrs total run time, for better accuracy I should have deducted 59k points from the 11.579,993 points of the 2 week period - which I didn't. So my results have a margin of error by about 0,5%.

The longer the observed period, the smaller the boundary errors will get (in the average).

August is a 31 day month. The upper bound my system could achieve is approx. 31 mio points. Given the trend as seen in the first 14 days, it will probably land somewhere in the 24-25 mio point range, which is 77 to 81% of peak performance. I will keep the single rig running for August, so the number should be a good approximation of an average performance with the current projects.

Based on the 615 watt power consumption, the energy efficiency is with 1310 ppd / watt excellent and significantly better than any GPU based solution. A dual Titan system with 450 watt would need 590.000 (average) ppd to match the i4P platform for energy efficiency.
 
Hi brilong. How could you overclock a 4p E5 rig? What tools are you using?

brilong said:
I overclock (all-cores 3176MHz according to i7z). 100% c0 state with temps up to 73C on hottest CPU. I'm using 1333 MHz ECC RAM.
Click to expand...
 
It is so easy to loose points :(

Two weeks ago I got a really good emergency power unit. While all the necessary cables to connect it to the hosue were still shipping, I had to fly to the US. It wasn't long that my wife called me that the whole district in the city is out of power. With the proper cables, and already properly set up (i.e filling all the tanks), it would have been so easy. It wasn't. To run the house on one of the smaller Honda generators, she had to unplug a lot of heaters, machines, and my servers to get below the max. performance limitation of the Honda generator. So the house were running on backup power, but not my servers.
While she fiddled with the smaller unit, my son came back and helped her to start the bigger unit and connected it with cables and adapters we had in the basement. The house were more or less back to normal, but my servers were still offline as I hadn't had time to config them for autostart which I wanted to do when setting up the whole generator thing. Came back over the weekend and restarted everything.

Anyway, 4 days folding lost :(
The system is now back to folding :)
Guess what the first WU was? P8101 :(
Second WU: P8104 :)

After the power outage, the shipment arrived with the new cables. This will allow me to run the house off the generator with basically no need to turn off anything vs. being connected to the grid.

The new extension cable connecting the generator and the main distribution panel has at least a decent plug
original.jpg


BTW, I only posted some lousy i4p photos in the past. Let me try to fix that.

Here is the SM board configured for I/O stuff:
1920x1280
original.jpg


Viewpoint like in Manhattan downtown :), 1920x1440
original.jpg


Better view on the 4P setup, 1920x1440
original.jpg


Andy
 
Yes a box folder a man behind my own heart :)

I see 1 minor problem though AndyE no standoffs, you may want to add some so you can get a little air flow around the board.

7wav.jpg
 
AndyE:

Request/Recommendation:

why don't you go ahead and create your own thread with just pictures of all your awesome hardware, then there will be a central area that people can go and drool over repeatedly.

yours truly,

bigteds
 
I remember when you used to drool over my hardware... where did I go wrong? :p

drool.jpg
 
Last edited:
tear said:
I remember when you used to drool over my hardware... where did I go wrong?
Click to expand...

I drooled so dam much that I just had to get a bunch of my own, I then ran out of room ond power outlets, I now have to be very mobile because when Grandma get's pissed they have to move quickly. That one is in the formal dining room which gets 86ed every time I get into trouble. It is all your fault, somebody made me to damn H]ard. :D
 
I believe so, they allow the heat to get away from the motherboard and heat is the biggest crippler of electronics, I have just always preferred to put standoff's on the motherboard, allot of motherboards come with standoffs at least the used to anyway.
 
^^ A lot of "modern" electronic components have the majority of their heat dissipation done through the PCB planes (power/ground). Solder slugs built into the chips allow the heat to be transferred into the board as a cheap heat-sink.
Standoffs will allow that heat that makes it through the board to get some air flow under there.
Highly recommended.
 
You must log in or register to reply here.
Back
Top