correlation between GPU utilisation and TF bit-range

mattmill30 · 2016-09-04, 18:50

Is anyone able to comprehensively explain to a level from which a calculation can be produced, the correlation between TF exponent size and bit-range (and process/sieve size?) against GPU saturation, in order to determine the performance limit of a GPU based upon it's published specs?

Would the focus be completion of classes at a certain rate for the transition overhead to allow graphical tasks to be undertaken in a timely fashion?

It would be useful to know the impact of different bit-ranges on the responsiveness of the hardware, so distribution of TF work can be intelligently assigned for utilisation of unused GPU power. e.g. Workstations.

LaurV · 2016-09-05, 06:20

Last versions(s) of mfaktx should occupy the GPU 100% when gpu-sieving is used, regardless of bit level and/or exponent, the only exceptions being the "extreme" cases, but those cases are also avoidable, either using the less-classes (420 classes) versions in one side of the chart, or launching multiple instances of mfaktx in the other side of the chart (or both).

The older versions or "extreme versions" of mfaktc (like TFing very small exponents, or the versions modified to sieve higher than 2p+1) which do CPU sieving only will only occupy as much from the GPU as the CPU is able to sieve. I remember that a 4-core CPU (i7-2600k) was not able to saturate two gtx580 cards more than 70% or so, the bottleneck being the CPU-sieving, when I was trying to TF exponents from 2k to 100k (in the ECM area).

James' tables are quite accurate approximations of the performance, for most cards there (exceptions are few "rare" cards with only few benchmarks).

Mark Rose · 2016-09-05, 13:00

For what it's worth, I've not found any configuration for acceptable desktop performance while using GPU sieving under Linux.

LaurV · 2016-09-06, 01:36

Quote:

Originally Posted by Mark Rose

for acceptable desktop performance

What do you mean by that? Is the desktop not enough performant when you use mfaktx, like it is not responsive, it has lag, etc., or you talk about mfaktx performance itself, like the cards are not 100% occupied, and the output of TF is slow?

Mark Rose · 2016-09-06, 01:45

Quote:

Originally Posted by LaurV

What do you mean by that? Is the desktop not enough performant when you use mfaktx, like it is not responsive, it has lag, etc., or you talk about mfaktx performance itself, like the cards are not 100% occupied, and the output of TF is slow?

The desktop becomes very laggy, from moving the mouse to scrolling. mfaktc performance is fine.

LaurV · 2016-09-06, 05:14

Oh, I got it, I was reading the OP's post totally wrong. I thought he has problems with the output, like he can't get the mfaktx to occupy the card 100% or the output is low. In fact, the problem is that the computer is not responsive. I can't help that under linux. On windoze, playing with the parameters in the ini file may slow mfaktc up to a level where it is not "sensed" by the normal graphic tasks, like playing a movie, or do some cad modeling. I usually don't play animation-intensive games. There is a discussion/tutorial here around, and even some batch files to launch mfaktc in different "hurry" modes.

mattmill30 · 2016-10-02, 17:20

Quote:

Originally Posted by mattmill30

Is anyone able to comprehensively explain to a level from which a calculation can be produced, the correlation between TF exponent size and bit-range (and process/sieve size?) against GPU saturation, in order to determine the performance limit of a GPU based upon it's published specs?

Would the focus be completion of classes at a certain rate for the transition overhead to allow graphical tasks to be undertaken in a timely fashion?

It would be useful to know the impact of different bit-ranges on the responsiveness of the hardware, so distribution of TF work can be intelligently assigned for utilisation of unused GPU power. e.g. Workstations.

Whilst this won't amount to my original request for a formula, would it be pseudomathmatically valid to ascertain the limit of efficiency for a particular piece of hardware by plotting a graph for each tf level of an exponent on a single piece of hardware, and concluding the efficiency limit where a spike in bit-level/time curve occurs?

Would it also be necessary to create a number of graphs against exponents of disproportionate sizes?

Uncwilly · 2016-10-03, 14:05

Quote:

Originally Posted by mattmill30

Would it also be necessary to create a number of graphs against exponents of disproportionate sizes?

James H has graphs like that on his website.
Here is a link to one: http://www.mersenne.ca/cudalucas.php?model=490

2016-09-04, 18:50	#1
mattmill30 Aug 2015 2×23 Posts	correlation between GPU utilisation and TF bit-range Is anyone able to comprehensively explain to a level from which a calculation can be produced, the correlation between TF exponent size and bit-range (and process/sieve size?) against GPU saturation, in order to determine the performance limit of a GPU based upon it's published specs? Would the focus be completion of classes at a certain rate for the transition overhead to allow graphical tasks to be undertaken in a timely fashion? It would be useful to know the impact of different bit-ranges on the responsiveness of the hardware, so distribution of TF work can be intelligently assigned for utilisation of unused GPU power. e.g. Workstations.

2016-09-05, 06:20	#2
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 10100000110011₂ Posts	Last versions(s) of mfaktx should occupy the GPU 100% when gpu-sieving is used, regardless of bit level and/or exponent, the only exceptions being the "extreme" cases, but those cases are also avoidable, either using the less-classes (420 classes) versions in one side of the chart, or launching multiple instances of mfaktx in the other side of the chart (or both). The older versions or "extreme versions" of mfaktc (like TFing very small exponents, or the versions modified to sieve higher than 2p+1) which do CPU sieving only will only occupy as much from the GPU as the CPU is able to sieve. I remember that a 4-core CPU (i7-2600k) was not able to saturate two gtx580 cards more than 70% or so, the bottleneck being the CPU-sieving, when I was trying to TF exponents from 2k to 100k (in the ECM area). James' tables are quite accurate approximations of the performance, for most cards there (exceptions are few "rare" cards with only few benchmarks). Last fiddled with by LaurV on 2016-09-05 at 06:26

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Spearman's rank correlation coefficient	Unregistered	Homework Help	2	2009-03-03 19:56
OK, how can we get a range now?	thechickenman	Lone Mersenne Hunters	4	2008-12-01 10:45
why 45% CPU utilisation?	optim	Software	1	2004-07-10 16:58
Available range for TF to 2^60?	edorajh	Lone Mersenne Hunters	2	2003-12-31 16:04
getting a range?	tom11784	Lone Mersenne Hunters	1	2003-08-29 18:56

2016-09-05, 13:00	#3
Mark Rose "/X\(‘-‘)/X\" Jan 2013 https://pedan.tech/ 6160₈ Posts	For what it's worth, I've not found any configuration for acceptable desktop performance while using GPU sieving under Linux.

2016-09-06, 05:14	#6
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 24063₈ Posts	Oh, I got it, I was reading the OP's post totally wrong. I thought he has problems with the output, like he can't get the mfaktx to occupy the card 100% or the output is low. In fact, the problem is that the computer is not responsive. I can't help that under linux. On windoze, playing with the parameters in the ini file may slow mfaktc up to a level where it is not "sensed" by the normal graphic tasks, like playing a movie, or do some cad modeling. I usually don't play animation-intensive games. There is a discussion/tutorial here around, and even some batch files to launch mfaktc in different "hurry" modes.