|
Message boards :
Number crunching :
Threading
| Author |
Message |
|---|
|
|
|
This was raised by Peter Hucker in the thread about splitting work units across GPUs. As that's now locked, here we are.
So am I running my 24 core work units slower than I should on my dual 12 core xeons? Would it even be clever enough to run two 12 core units on seperate CPUs?
yes, although even 12 core units may not be fastest, you need to test it.
I've seen people running tasks on a threadripper using 24 cores that end up slower than my 3700x running on 8.
Today I found my wingman using a 3950x who's running 16 SoB tasks concurrently.
This is the result
Around 110000s/task.
around 3x slower then the 35-40ks/task I get on my 3950x when running 2 at a time with 8 threads each.
I've PM'd the user to point this out, I'm hoping he understands the message.
There should really be a link in the options on these big tasks to an info page explaining how to check you're getting the best use out of your cpu.
| |
|
|
|
This was raised by Peter Hucker in the thread about splitting work units across GPUs. As that's now locked, here we are.
So am I running my 24 core work units slower than I should on my dual 12 core xeons? Would it even be clever enough to run two 12 core units on seperate CPUs?
yes, although even 12 core units may not be fastest, you need to test it.
I've seen people running tasks on a threadripper using 24 cores that end up slower than my 3700x running on 8.
Today I found my wingman using a 3950x who's running 16 SoB tasks concurrently.
This is the result
Around 110000s/task.
around 3x slower then the 35-40ks/task I get on my 3950x when running 2 at a time with 8 threads each.
I've PM'd the user to point this out, I'm hoping he understands the message.
There should really be a link in the options on these big tasks to an info page explaining how to check you're getting the best use out of your cpu.
Seti did a survey a while back and found out that less than 10% of crunchers ever use the message boards at all, a few have problems and ask but can't figure out the answer so they just move on to something else or quit crunching altogether. MOST people use the default settings never knowing that tweaking them can make their system run much better. ie how many people have to be told to change their cache settings from 10 days to much less here at PrimeGrid so they can have a chance at being #1 in finding Primes?
According to BoincStats PrimeGrid has Active users 2,985, there is no way all of those people are on these message boards. Even those with multiple pc's may not come here, Hosts 45,202 | |
|
|
|
|
That people don't bother with the message boards is exactly why I'm suggesting a notice & link on the task selection screen would be a good idea. | |
|
|
|
|
As was pointed out on the other thread...different people have different crunching goals. And these goals may be achieved better if run with a non optimal cpu/gpu strategy. In other words, one persons "best use" in not necessarily another persons "best use."
____________
Werinbert is not prime... or PRPnet keeps telling me so.
Badge score: 1x1 + 12x3 + 1x4 + 1x5 + 1x6 + 2x7 + 1x8 + 1x10 = 84 | |
|
 Bur Volunteer tester
Send message
Joined: 25 Feb 20
Posts: 332
ID: 1241833
Credit: 22,611,276
RAC: 0
               
|
Today I found my wingman using a 3950x who's running 16 SoB tasks concurrently.
This is the result
Around 110000s/task.
around 3x slower then the 35-40ks/task I get on my 3950x when running 2 at a time with 8 threads each.
3x slower, but he's running 8x more tasks. So, I'm not sure, but doesn't that mean his throuput is actually higher than yours? He does 16 tasks in 110000 s, or on average one task every 6875 s.
You only finish one task per 17000-20000 s. This might have made sense when being 1st meant something, but now it just means less chance of finding a prime.
That difference seems a bit extreme, so maybe I got the numbers you posted wrong?
____________
Primes: 1281979 & 12+8+1979 & 1+2+8+1+9+7+9 & 1^2+2^2+8^2+1^2+9^2+7^2+9^2 & 12*8+19*79 & 12^8-1979 & 1281979 + 4 (cousin prime) | |
|
|
|
Today I found my wingman using a 3950x who's running 16 SoB tasks concurrently.
This is the result
Around 110000s/task.
around 3x slower then the 35-40ks/task I get on my 3950x when running 2 at a time with 8 threads each.
3x slower, but he's running 8x more tasks. So, I'm not sure, but doesn't that mean his throuput is actually higher than yours? He does 16 tasks in 110000 s, or on average one task every 6875 s.
You only finish one task per 17000-20000 s. This might have made sense when being 1st meant something, but now it just means less chance of finding a prime.
That difference seems a bit extreme, so maybe I got the numbers you posted wrong?
This is my assessment as well. With everyone being first, the ideal situation is to optimize around the throughput of tasks, rather than the throughput of firsts like when it used to be a race.
____________
 | |
|
|
|
Today I found my wingman using a 3950x who's running 16 SoB tasks concurrently.
This is the result
Around 110000s/task.
around 3x slower then the 35-40ks/task I get on my 3950x when running 2 at a time with 8 threads each.
3x slower, but he's running 8x more tasks. So, I'm not sure, but doesn't that mean his throuput is actually higher than yours? He does 16 tasks in 110000 s, or on average one task every 6875 s.
You only finish one task per 17000-20000 s. This might have made sense when being 1st meant something, but now it just means less chance of finding a prime.
That difference seems a bit extreme, so maybe I got the numbers you posted wrong?
Your maths is flawed.
He's doing 16 tasks in 1700000 seconds or 110000s/task.
I do 2 tasks in 70000 seconds or 35000s/task.
Prime95 will verify this difference if you run the benchmarks. 1 task will use the entire cache on a ryzen ccd so 16 tasks will be running mainly in ram. | |
|
Post to thread
Message boards :
Number crunching :
Threading |
