SMP Affinity Changer

[compuguy1088]

Works ok, here.

Just checked again, and the site is up on my end as well.

[rilian]

Hello, compuguy1088, thank you!

that could be temporary server outage or smth, dont pay attention

[britgeezer]

I have been using Affinity Changer for a while now with 2x SMP 5.91 on my Quad Q6600 XP SP3 systems - Thank you

Now I'm also running GPU2 and need to consider core (3) should be dedicated to GPU2, I'm disappointed with output on SMP CPU with 6.22 beta running 1x SMP and in fact found that limiting the SMP to cores 0 & 1 increased output.

I know from your FAC that SMP_FC is not intended for 1x SMP, but its surprising that 3 cores gives less output than 2 set with affinity manually under Windows Task manager.

Is this something that the next version will consider, indeed, should I try 2xSMP with only 3 cores?

BG

[toTOW]

Affinity changer isn't working very well when you run a GPU client ... you should consider not using it with SMP + GPU setups.

[britgeezer]

Yes I had thought the same and temporaly removed the program.
Does anyone have an opinion on the value of setting affinity manually to 3 or even perhaps 2 cores for SMP 6.22 when also using GPU2?
GPU2 defines core 3 automatically.

[rilian]

Works good at crunchers from my team with following:

2хSMP + ACh + 2xGPU clients

...

C2Quad Q6600 (3600МГц) / MEM 2x1024 GoodRam DDRII 900 (1125МГц) / Abit IP35 Pro / 2xWinSMP (ACh) +
GF 8800GTS 512MB (761/1902/2000) + GF 9600GSO 384MB (700/1901/1600)

gives around 14000 PPD

[britgeezer]

Works good at crunchers from my team with following:

2хSMP + ACh + 2xGPU clients

...

C2Quad Q6600 (3600МГц) / MEM 2x1024 GoodRam DDRII 900 (1125МГц) / Abit IP35 Pro / 2xWinSMP (ACh) +
GF 8800GTS 512MB (761/1902/2000) + GF 9600GSO 384MB (700/1901/1600)

gives around 14000 PPD

Rilian,

How are the sytems setup do you do anything to the standard core 3 setting of GPU2?

[rilian]

will this help?

from our forum

http://distributed.org.ua/forum/index.p ... entry29222

Code: Select all

[settings]
username=Vorchun
team=2164
passkey=хххххххххххххххххххххххххххххххх
asknet=no
machineid=3
bigpackets=big
extra_parms=-gpu 0
local=

[http]
active=no
host=localhost
port=8080
usereg=no
proxy_name=
proxy_passwd=

[core]
priority=100
cpuusage=100
disableassembly=no
nocpulock=0
checkpoint=30
ignoredeadlines=yes

[power]
battery=no

[clienttype]
memory=1024
type=3

Another one:
http://distributed.org.ua/forum/index.p ... entry27856

Optimal params for Quad and two GPU

2GPU, 2SMP, ACh
Do not create NV_FAH_CPU_AFFINITY in registry, core of GPU1 assigns itself to CPU_core_4, core of GPU2 assigns itself to CPU_core_3. PPD for SPM drops on 12-15% (!)

GPU params: priority=100, cpuusage=100

===========
For Core Duo and one GPU (C2D E8200 + 8800GT) - 1GPU client and 1 SMP client

Create NV_FAH_CPU_AFFINITY = 0 so FahCore_11.exe will assign to both CPU cores. Use ACh
In this case GPU_PPD = 4400, SMP_PPD=1900 (Cpu at 3000MHz)

[gwildperson]

Now I'm also running GPU2 and need to consider core (3) should be dedicated to GPU2, I'm disappointed with output on SMP CPU with 6.22 beta running 1x SMP and in fact found that limiting the SMP to cores 0 & 1 increased output.

Did you actually measure this? With nothing except one 6.22beta client running, what are the frame times if you manually set all four fahcore_a1's to
* Cores 0 and 1
* Cores 0 1 and 2
* Cores 0 1 2 and 3

From what I read, fahcore_a1 doesn't scale very well and will soon be replaced by fahcore_a2 so I'm not sure of the long-term value of these measurments. In any case, I find it hard to believe that the first one is faster than the third, but don't have any solid information.

Maybe by "increased output" are you also considering other clients are doing simultaneously. I have no doubt that the third one is less that twice as fast as the first one. Because of that, two clients (cores 0,1 and cores 2,3) does more work than just one of them, which is what "does not scale well" means.

[britgeezer]

Now I'm also running GPU2 and need to consider core (3) should be dedicated to GPU2, I'm disappointed with output on SMP CPU with 6.22 beta running 1x SMP and in fact found that limiting the SMP to cores 0 & 1 increased output.

Did you actually measure this? With nothing except one 6.22beta client running, what are the frame times if you manually set all four fahcore_a1's to
* Cores 0 and 1
* Cores 0 1 and 2
* Cores 0 1 2 and 3

From what I read, fahcore_a1 doesn't scale very well and will soon be replaced by fahcore_a2 so I'm not sure of the long-term value of these measurments. In any case, I find it hard to believe that the first one is faster than the third, but don't have any solid information.

Maybe by "increased output" are you also considering other clients are doing simultaneously. I have no doubt that the third one is less that twice as fast as the first one. Because of that, two clients (cores 0,1 and cores 2,3) does more work than just one of them, which is what "does not scale well" means.

Simply based on FHmon, on the suggestion of a forum friend, I tried it and noticed no difference to SMP output and and increase on GPU2. You can imagine that any use of that PC benefited from the spare core, just a shame to see it "wasted".

[britgeezer]

will this help?

from our forum

http://distributed.org.ua/forum/index.p ... entry29222

Code: Select all

[settings]
username=Vorchun
team=2164
passkey=хххххххххххххххххххххххххххххххх
asknet=no
machineid=3
bigpackets=big
extra_parms=-gpu 0
local=

[http]
active=no
host=localhost
port=8080
usereg=no
proxy_name=
proxy_passwd=

[core]
priority=100
cpuusage=100
disableassembly=no
nocpulock=0
checkpoint=30
ignoredeadlines=yes

[power]
battery=no

[clienttype]
memory=1024
type=3

Another one:
http://distributed.org.ua/forum/index.p ... entry27856

Optimal params for Quad and two GPU

2GPU, 2SMP, ACh
Do not create NV_FAH_CPU_AFFINITY in registry, core of GPU1 assigns itself to CPU_core_4, core of GPU2 assigns itself to CPU_core_3. PPD for SPM drops on 12-15% (!)

GPU params: priority=100, cpuusage=100

===========
For Core Duo and one GPU (C2D E8200 + 8800GT) - 1GPU client and 1 SMP client

Create NV_FAH_CPU_AFFINITY = 0 so FahCore_11.exe will assign to both CPU cores. Use ACh
In this case GPU_PPD = 4400, SMP_PPD=1900 (Cpu at 3000MHz)

Which version of SMP please, I note there was no need to use -SMP and instead I see - GPU 0 or is this thre config for GPUx?

[rilian]

i'll ask what exactly version used later

[rilian]

britgeezer

we use version 5.91

6.22 is not very stable

--

now, on the above config,

With 2xGPU running, SMP1 gives 2016 PPD, SMP2 gives 2045 PPD
Without 2xGPU running SMP clients will give both 2300 PPD (on p2653 WUs)

[rilian]

Note, Fah SMP 6.22 has only one PROCESS with 4 threads, while early versions (5.91) run 4 PROCESSES each with one thread. Process may be assigned to some core, thread - no (at least by outer program). That is why using affinity changer does not make sense and you do not get additional 10% speedup.

i also do not know whether this new fah client assign affinities correctly (i would guess "no")

[britgeezer]

Note, Fah SMP 6.22 has only one PROCESS with 4 threads, while early versions (5.91) run 4 PROCESSES each with one thread. Process may be assigned to some core, thread - no (at least by outer program). That is why using affinity changer does not make sense and you do not get additional 10% speedup.

i also do not know whether this new fah client assign affinities correctly (i would guess "no")