mfaktc: a CUDA program for Mersenne prefactoring - Page 110

fivemack · 2011-09-15, 13:23

Quote:

They both run with the same Sieve Prime setting. (And auto-adjusting is on.)
This does not occur in Windows. With the same affinities and Prime95 running 1-6, I still get around 165-170 with mfaktc.

The CPU numbering isn't necessarily the same in Windows and Linux; it's entirely possible that Windows has the pairing of hyperthreads as (12)(34)(56)(78) whilst, on my i7/920, I know that Linux has (18)(27)(35)(46).

Could you put up the result from 'grep apicid /proc/cpuinfo' ?

Dubslow · 2011-09-16, 05:25

Hmm... Since I booted back to Windows to get back to the really fast rates, suddenly Windows only gets 75-80M/s... I didn't change anything!!!! What the hell?!?!? :(
Also, in Windows, it increases Sieve Primes massively until the average wait is down to a decent number; that doesn't happen in Linux with 75-80M/s rate. So confused x(
EDIT: Let me clarify. Currently, mfaktc-lin is running ~90M/s, Sieve Primes=5000 (adjust=1) and avg. wait is <100 us. Most recently in Windows, mfaktc was running 90M/s rate, 5000 SP, and 3000+ us avg. wait at start, which would shift to 75M/s, 15000+ SP, and 300-400 us avg wait after enough classes completed. I don't know why/how this behavior changed from the 170M/s +, as far as I know I didn't change anything.

Imma boot to Linux now and run that command. On the other hand, I remember testing it, and it seemed pretty clear to me that it was (12)(34)(56)(78), even though MPrime detects (15)(26)(37)(48). Huh.

This computer hates me.

EDIT:

Code:

apicid		: 0
initial apicid	: 0
apicid		: 2
initial apicid	: 2
apicid		: 4
initial apicid	: 4
apicid		: 6
initial apicid	: 6
apicid		: 1
initial apicid	: 1
apicid		: 3
initial apicid	: 3
apicid		: 5
initial apicid	: 5
apicid		: 7
initial apicid	: 7

Could you please explain what exactly APIC id is?

Also also, each OS seems incapable of reading the checkpoint files of the other. Should I be converting those too?

TheJudger · 2011-09-16, 10:25

Hi,

Quote:

Originally Posted by Dubslow

Also also, each OS seems incapable of reading the checkpoint files of the other. Should I be converting those too?

this is not a CR/CRLF/LF problem. For safety reason mfaktc refuses to load checkpoint files from other versions than itself. Windows and Linux versions have a different version string.

I might change this in the future but not now.

Oliver

fivemack · 2011-09-16, 10:52

Quote:

Originally Posted by Dubslow

Code:

apicid		: 0
initial apicid	: 0
apicid		: 2
initial apicid	: 2
apicid		: 4
initial apicid	: 4
apicid		: 6
initial apicid	: 6
apicid		: 1
initial apicid	: 1
apicid		: 3
initial apicid	: 3
apicid		: 5
initial apicid	: 5
apicid		: 7
initial apicid	: 7

Could you please explain what exactly APIC id is?

It's a hardware identifier for the processors; the two hyperthreads on processor n have ID 2n and 2n+1, so what you're seeing is (15)(26)(37)(48) as mprime has picked up. I've no idea how Linux assigns the processor numbers that it uses ... I'm reasonably trusting that, on an idle single-chip machine, a job running four threads will get them on four distinct processors rather than having two run in hyperthreads on the same processor, but on more complicated systems it doesn't seem to work as perfectly as I would hope.

Dubslow · 2011-09-16, 14:59

Urrgggg....
Now Windows is back to 165-170M/s again and I still haven't changed anything...
In fact, I haven't even shut it down in any way since the last time it was at 75M/s...
Thanks fivemack, I'll look into modifying my MPrime settings.

Dubslow · 2011-09-16, 21:14

RRRRRRRAAAAAAAAAGGGGGGGGGEEEEEEEEEE

mfaktc was going at 165M/s. I closed and immediately reopened it, and then it went down to 70M/s again. WHAT CHANGEDa
sdfnva'jub;gvsopofIZJX
IOHNS
io'j;JUB;;kbguxhdvc

TheJudger · 2011-09-16, 21:51

Dubslow: just to be sure

precompiled binary or selfcompiled?
settings in mfaktc.ini (if not default)?

Did you check if you GPU runs in highest performance mode? On Windows you can use GPU-Z to check the actuall GPU speed (Tab: Sensors). On Linux the you can start nvidia-settings to view the actual performance mode.

Oliver

Dubslow · 2011-09-17, 09:09

Pre compiled, 1.7. I couldn't compile anything in Windows for the life of me. I think I had 4 streams, SP=5000. Note: normally I keep SPAdjust=1, but when it's at 90M/s, the avg. wait skyrockets, so the SP goes up, which winds up reducing the avg. rate to 75M/s, so I decided to keep SP stuck to 5000 until I can get back up to previous level, i.e. 165-170M/s.

Code:

SievePrimes=5000
# Set this to 1 to enable automatically adjustments of SievePrimes during

# runtime based on the "average wait times".


SievePrimesAdjust=0


# Set the number of CUDA streams / data sets used by mfaktc.

# NumStreams must be >= 1. In this case mfaktc can process one stream /

# data set on the GPU while the GPU can preprocess the other one. When

# NumStreams is >= 2 than the time needed to upload (CPU->GPU transfer)

# the data sets can be hidden (if the hardware supports it (most GPUs are

# supporting this)).

# On Linux systems 2 or 3 seems a good numbers. There are comments that

# Windows systems need a greater number of streams.

# A greater number increases the memory consumed by mfaktc (host and GPU

# memory). The current limit for the number of streams is 10!


NumStreams=4
# Set the number of data sets which can be preprocessed on CPU. This allows

# to tolerate more jitter on runtime of preprocessing and GPU stream

# runtime.


CPUStreams=4
# The GridSize affects the number of threads per grid.

# Depending on the number of multiprocessors of your GPU, too, the

# automatic parameter threads per grid is set to:

#   GridSize = 0:  65536 < threads per grid <=  131072

#   GridSize = 1: 131072 < threads per grid <=  262144

#   GridSize = 2: 262144 < threads per grid <=  524288

#   GridSize = 3: 524288 < threads per grid <= 1048576 (default)

# A smaller GridSize has more overhead than a bigger GridSize for long

# running jobs. For really small jobs there can be a small benefit on

# computation time if the GridSize is small. A smaller GridSize directly

# reduces the runtime per kernel launch and might result in a better

# interactivity one older GPUs.


GridSize=3
# WorkFile: the name of the file which contains the factoring assignments.

# e.g.

# worktodo.ini (Prime95 v24 and earlier)

# worktodo.txt (Prime95 v25 and newer)


WorkFile=worktodo.txt



# Checkpoints = 0: disable checkpoints

# Checkpoints = 1: enable checkpoints

# Checkpoints are needed for resume capability, after a class is finished a

# checkpoint file is written. When mfaktc is interrupted during the run and

# restarted later it will begin at the last processed class.


Checkpoints=1



# Allow to split an assignment into multiple bit ranges.

# 0 = disabled

# 1 = enabled

# Enabled Stages make only sense when StopAfterFactor is 1 or 2.


Stages=0


# possible values for StopAfterFactor:

# 0: Do not stop the current assignment after a factor was found.

# 1: When a factor was found for the current assignment stop after the

#    current bitlevel. This makes only sense when Stages is enabled.

# 2: When a factor was found for the current assignment stop after the

#    current class.


StopAfterFactor=2


# possible values for PrintMode:

# 0: print a new line for each finished class

# 1: overwrite the current line (more compact output)


PrintMode=1


# allow the CPU to sleep if nothing can be preprocessed?
# 0: Do not sleep if the CPU must wait for the GPU

# 1: The CPU can sleep for a short time if it has to wait for the 
GPU

AllowSleep=0

The frustrating thing is that I didn't change anything.

I do know it's not related to GPU clock rate. (I use MSI Afterburner in Windows.) When it's at 165M/s, changing the clock rate directly affects the avg. rate. I just tested the 90M/s half rate now: I verified it was somewhat OC'ed. I killed mfaktc, reset the clock to stock, restarted mfaktc, got exactly the same avg. rate (90M/s), then upped the rate back to slight OC (with mfaktc running this time) and the avg. rate didn't budge. (I have previously tested that avg. rate changes with clock rate changes, while mfaktc is running.) This seems to indicate to me that it's something funky with the program. GPU load is 86-87%, same as it was when running at 165M/s (I got that reading sometime last week.)

EDIT: Oh my god Windows/Linux I swear there's no extra lines in mfaktc.cfg or whatever it's called. lol, that's what you guys said :P:)

Dubslow · 2011-09-17, 09:21

Hmm, just rebooted, now it's back to 170M/s, though as I said a few posts back, it's not just the rebooting...
How can I restart the graphics driver without rebooting?
Also, GPU-Z now shows 75-76% load, not the 85% I said earlier.

Ralf Recker · 2011-09-17, 09:50

Quote:

Originally Posted by Dubslow

I do know it's not related to GPU clock rate. (I use MSI Afterburner in Windows.) When it's at 165M/s, changing the clock rate directly affects the avg. rate. I just tested the 90M/s half rate now: I verified it was somewhat OC'ed. I killed mfaktc, reset the clock to stock, restarted mfaktc, got exactly the same avg. rate (90M/s), then upped the rate back to slight OC (with mfaktc running this time) and the avg. rate didn't budge.

This is probably the driver downclock bug.

Karl M Johnson · 2011-09-17, 14:34

Quote:

Originally Posted by Ralf Recker

This is probably the driver downclock bug.

Yes, exiting from CUDA applications incorrectly causes top perf. levels clocks to be stuck at 410 Mhz core.

2011-09-16, 05:25	#1201
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Hmm... Since I booted back to Windows to get back to the really fast rates, suddenly Windows only gets 75-80M/s... I didn't change anything!!!! What the hell?!?!? :( Also, in Windows, it increases Sieve Primes massively until the average wait is down to a decent number; that doesn't happen in Linux with 75-80M/s rate. So confused x( EDIT: Let me clarify. Currently, mfaktc-lin is running ~90M/s, Sieve Primes=5000 (adjust=1) and avg. wait is <100 us. Most recently in Windows, mfaktc was running 90M/s rate, 5000 SP, and 3000+ us avg. wait at start, which would shift to 75M/s, 15000+ SP, and 300-400 us avg wait after enough classes completed. I don't know why/how this behavior changed from the 170M/s +, as far as I know I didn't change anything. Imma boot to Linux now and run that command. On the other hand, I remember testing it, and it seemed pretty clear to me that it was (12)(34)(56)(78), even though MPrime detects (15)(26)(37)(48). Huh. This computer hates me. EDIT: Code: apicid : 0 initial apicid : 0 apicid : 2 initial apicid : 2 apicid : 4 initial apicid : 4 apicid : 6 initial apicid : 6 apicid : 1 initial apicid : 1 apicid : 3 initial apicid : 3 apicid : 5 initial apicid : 5 apicid : 7 initial apicid : 7 Could you please explain what exactly APIC id is? Also also, each OS seems incapable of reading the checkpoint files of the other. Should I be converting those too? Last fiddled with by Dubslow on 2011-09-16 at 05:45

2011-09-16, 21:51	#1206
TheJudger "Oliver" Mar 2005 Germany 457₁₆ Posts	Dubslow: just to be sure precompiled binary or selfcompiled? settings in mfaktc.ini (if not default)? Did you check if you GPU runs in highest performance mode? On Windows you can use GPU-Z to check the actuall GPU speed (Tab: Sensors). On Linux the you can start nvidia-settings to view the actual performance mode. Oliver

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1676	2021-06-30 21:23
The P-1 factoring CUDA program	firejuggler	GPU Computing	753	2020-12-12 18:07
gr-mfaktc: a CUDA program for generalized repunits prefactoring	MrRepunit	GPU Computing	32	2020-11-11 19:56
mfaktc 0.21 - CUDA runtime wrong	keisentraut	Software	2	2020-08-18 07:03
World's second-dumbest CUDA program	fivemack	Programming	112	2015-02-12 22:51

2011-09-16, 14:59	#1204
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 1C35₁₆ Posts	Urrgggg.... Now Windows is back to 165-170M/s again and I still haven't changed anything... In fact, I haven't even shut it down in any way since the last time it was at 75M/s... Thanks fivemack, I'll look into modifying my MPrime settings.

2011-09-16, 21:14	#1205
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	RRRRRRRAAAAAAAAAGGGGGGGGGEEEEEEEEEE mfaktc was going at 165M/s. I closed and immediately reopened it, and then it went down to 70M/s again. WHAT CHANGEDa sdfnva'jub;gvsopofIZJX IOHNS io'j;JUB;;kbguxhdvc

2011-09-17, 09:21	#1208
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Hmm, just rebooted, now it's back to 170M/s, though as I said a few posts back, it's not just the rebooting... How can I restart the graphics driver without rebooting? Also, GPU-Z now shows 75-76% load, not the 85% I said earlier.