GPU Noob-Experiences/Questions - Page 10

garo · 2011-10-29, 12:28

Thanks Oliver. I didn't mean to imply that you should code a new version. Your contribution to the project has already been very valuable. GPU-TF is the best thing to have happened to the project since the P4 and v21. And you have contributed a lot of crunching too. Right from the days when you gave the old TPR a scare with your march up the TF stats. Old times!

Christenson · 2011-10-30, 14:45

On the AMD 6-core Phenom II, I was able to tell X-windows to run the on-board ATI/AMD video at the same time as the NVidia GPU. But I don't run Photoshop, HD movies, or games on that box at all....and lately, the box itself is sick, so I need to post a "computer diet -- need help!" thread.

TheJudger · 2011-10-30, 16:55

Hi Garo,

Quote:

Originally Posted by garo

Thanks Oliver. I didn't mean to imply that you should code a new version. Your contribution to the project has already been very valuable.

I wouldn't be a big deal to write another kernel which takes benefit (can handle) ILP. But I don't think it will be faster (at least on current hardware). Main issue with ILP are the add/sub with carry which are used quiet often in current kernels. Without them I need to do soft carries and I can't use the full 32 bit per variable for data. This means alot additional shifts, adds and bitmasks. Current kernels do not run totally bad on CC 2.1. So the hypothetical kernels tuned for ILP can better utilize the GPU cores but might be slower anyway.

Quote:

Originally Posted by garo

And you have contributed a lot of crunching too. Right from the days when you gave the old TPR a scare with your march up the TF stats. Old times!

How many years ago?
Seems that my password for the arstechnica forum does not work anymore? I wanted to reply on Paulie that I won't bet on Mac support for mfaktc.

Oliver

garo · 2011-10-30, 17:31

I think it was at least 7 years ago, maybe more. Doesn't Ars have a lost password link. I'll post that response for Paulie if you like.

davieddy · 2011-10-31, 06:11

Quote:

Originally Posted by garo

I think it was at least 7 years ago, maybe more. Doesn't Ars have a lost password link. I'll post that response for Paulie if you like.

Ars longa vita brevis

Wizzard · 2011-10-31, 23:27

I would like to ask, how to set mfaktc to slow down my computer at least as possible. (my hw: Core 2 Duo E6600@2880 MHz, 4 GB DDR2, GeForce 250 GTS 512 MB, running in 64-bit Linux. I have set it to NumStreams=1, CPUStreams=1, GridSize=1, is that OK?

Christenson · 2011-11-01, 01:14

Use your OS to set the priority of mfaktc as low as possible. And tell it to fix SievePrimes at 10000 (do relatively less sieving out of candidates on the CPU). The other settings you mentioned are also OK.

All this comes at a price in how much work mfaktc does, of course...but the 250 is a relatively low end card these days. You could also see if your card supports CUDALucas and have it do that..that will also slow down the card and therefore the impact on your CPU.

kladner · 2011-11-01, 01:24

Quote:

Originally Posted by Wizzard

I would like to ask, how to set mfaktc to slow down my computer at least as possible. (my hw: Core 2 Duo E6600@2880 MHz, 4 GB DDR2, GeForce 250 GTS 512 MB, running in 64-bit Linux. I have set it to NumStreams=1, CPUStreams=1, GridSize=1, is that OK?

Something I have been trying in the last couple of days is to set "AllowSleep=1" in mfaktc.ini. It's the last line in the file. It leaves the CPU at a little less than 100%. The difference in mfaktc performance seems to be minimal if you're not doing anything else on the machine. But other apps seem a bit more responsive.

I have never tried setting NumStreams, CPUStreams, or GridSize that low. I have those back at 3,3,3, now. I was running them at 2,3,2 for a while, but it did really slow down mfaktc. I'm not sure it helped that much with other apps.

All this is on a Phenom II 1090T x6, GTX 460, 8GB RAM, with 2 or 3 instances of mfaktc. Right now I'm only running 2 because I'm trying to finish up some partial LL assignments on one core.

EDIT: This is with Win7-64. The priority setting approach I use would not apply in your case. But as Christenson said, CUDALucas has much less CPU load than mfaktc. I bow to his experience regarding the Streams and Gridsize settings.

Christenson · 2011-11-01, 02:33

Don't bow to my experience here..I haven't fooled with numStreams, CPUStreams, etc that much..one core of that 1090T saturates a GTX440 card, which is what I play with....go out and try it...that's what I meant. I forgot about AllowSleep (!) it's a good idea too.

kladner · 2011-11-01, 02:44

Quote:

Originally Posted by Christenson

Don't bow to my experience here..I haven't fooled with numStreams, CPUStreams, etc that much..one core of that 1090T saturates a GTX440 card, which is what I play with....go out and try it...that's what I meant. I forgot about AllowSleep (!) it's a good idea too.

OK! Go out and try it. That's a good philosophy. You never know what might turn up in a particular setup, with particular personal tolerances.

With the GTX 460, 2 cores of the 1090T keep it pretty busy, though I get a decent increase in productivity with 3 instances/3 cores. As I said, I'm only doing 2 right now so that I have one CPU core to clean up partial LL assignments. The other three are doing P-1.

Wizzard · 2011-11-01, 09:06

Thank you very much guys, it helped me a lot. I would also like to try some LL testing on my GPU, but cannot find that precompiled GPULucas for Linux64...

2011-10-31, 23:27	#105
Wizzard Feb 2011 Bratislava 3³ Posts	I would like to ask, how to set mfaktc to slow down my computer at least as possible. (my hw: Core 2 Duo E6600@2880 MHz, 4 GB DDR2, GeForce 250 GTS 512 MB, running in 64-bit Linux. I have set it to NumStreams=1, CPUStreams=1, GridSize=1, is that OK? Last fiddled with by Wizzard on 2011-10-31 at 23:41

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Some noob questions about differential equation ?	awholenumber	Math	7	2017-06-18 07:25
Noob Question: What to use	bozocv	Msieve	36	2015-12-31 00:12
2 Noob questions	FlightTribe	Information & Answers	13	2012-11-28 19:57
noob poly questions	sleigher	Msieve	56	2011-11-17 11:37
Noob question	xago666	Information & Answers	3	2008-03-11 01:35

2011-10-29, 12:28	#100
garo Aug 2002 Termonfeckin, IE 5320₈ Posts	Thanks Oliver. I didn't mean to imply that you should code a new version. Your contribution to the project has already been very valuable. GPU-TF is the best thing to have happened to the project since the P4 and v21. And you have contributed a lot of crunching too. Right from the days when you gave the old TPR a scare with your march up the TF stats. Old times!

2011-10-30, 14:45	#101
Christenson Dec 2010 Monticello 1795₁₀ Posts	On the AMD 6-core Phenom II, I was able to tell X-windows to run the on-board ATI/AMD video at the same time as the NVidia GPU. But I don't run Photoshop, HD movies, or games on that box at all....and lately, the box itself is sick, so I need to post a "computer diet -- need help!" thread.

2011-10-30, 17:31	#103
garo Aug 2002 Termonfeckin, IE 2⁴×173 Posts	I think it was at least 7 years ago, maybe more. Doesn't Ars have a lost password link. I'll post that response for Paulie if you like.

2011-11-01, 01:14	#106
Christenson Dec 2010 Monticello 703₁₆ Posts	Use your OS to set the priority of mfaktc as low as possible. And tell it to fix SievePrimes at 10000 (do relatively less sieving out of candidates on the CPU). The other settings you mentioned are also OK. All this comes at a price in how much work mfaktc does, of course...but the 250 is a relatively low end card these days. You could also see if your card supports CUDALucas and have it do that..that will also slow down the card and therefore the impact on your CPU.

2011-11-01, 02:33	#108
Christenson Dec 2010 Monticello 5·359 Posts	Don't bow to my experience here..I haven't fooled with numStreams, CPUStreams, etc that much..one core of that 1090T saturates a GTX440 card, which is what I play with....go out and try it...that's what I meant. I forgot about AllowSleep (!) it's a good idea too.

2011-11-01, 09:06	#110
Wizzard Feb 2011 Bratislava 3³ Posts	Thank you very much guys, it helped me a lot. I would also like to try some LL testing on my GPU, but cannot find that precompiled GPULucas for Linux64...