More cores, worse performance? Why?

Message boards : Number crunching : More cores, worse performance? Why?
Message board moderation

To post messages, you must log in.

AuthorMessage
6BQ5

Send message
Joined: 7 Dec 18
Posts: 29
Credit: 12,725,636
RAC: 357
United States
Message 2036603 - Posted: 7 Mar 2020, 23:02:26 UTC

Not knowing that SETI@Home was planning to go into hibernation :( I got myself a HP ProLiant ML350p Gen 8 server to crunch some numbers. It has two E5-2620 v2 processors with 8 cores each. That makes for 16 physical cores and 32 logical cores with HyperThreading.

I installed BOINC, added SETI@Home to the project list, and set the maximum CPU allocation to 97%. Allocating 97% of my CPUs limits the number of tasks to 31. That should leave at least one logical core free for general purpose "stuff" like navigating the computer.

This server is nicknamed "The-Beast" and it's profile is available at the link below. Hopefully the link is publicly accessible.

https://setiathome.berkeley.edu/show_host_detail.php?hostid=8913028

I'm really surprised at how low the average credit value is. It's barely in the 1500 - 2000 range. Compare this to my gaming rig nicknamed "C1-PC".

https://setiathome.berkeley.edu/show_host_detail.php?hostid=8630889

That machine has an i9-9900KS which offers 16 logical cores. I've configured it to use 15 logical cores. My gaming rig can easily get 25,000 - 30,000 average credit depending on the specific tasks it is crunching through.

I understand my C1-PC is more modern and powerful than The-Beast but how can there be such a huge disparity between the two machines? The C1-PC has a GeForce 1080 card. Does the GPU card really make that much of a difference?

Thoughts?

-=- Boris
ID: 2036603 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13797
Credit: 208,696,464
RAC: 304
Australia
Message 2036607 - Posted: 7 Mar 2020, 23:13:29 UTC - in response to Message 2036603.  
Last modified: 7 Mar 2020, 23:26:15 UTC

I'm really surprised at how low the average credit value is. It's barely in the 1500 - 2000 range. Compare this to my gaming rig nicknamed "C1-PC".
My gaming rig can easily get 25,000 - 30,000 average credit depending on the specific tasks it is crunching through.
Thoughts?
It takes 6-8 weeks (if there are no server issues, 8-12 weeks if there are) for RAC to end up around it's nominal value.
Since your new system has been running for only 2 days, there is no way in hell it will get to reach it's nominal RAC before Seti stops issuing new work, even if there are no more server issues between now & then.


The C1-PC has a GeForce 1080 card. Does the GPU card really make that much of a difference?
Yes.
eg. my Linux system just retired, a WU that takes 1hr 45min on the CPU is done by an RTX 2060 Super in 1 min 15 secs.
Roughly 84 times the output of 1 CPU core. Even with 12 CPU cores, it's still a long way behind the output of that 1 mid range video card. A high-end video card will have more than double the output of mine.


Edit- oh, and clock speed makes a big difference too. A system with half the clock speed of another needs twice as many cores to produce the same amount of work.
And that's what you've got, a system with twice as many cores, running at half the clock speed (probably less than half due to the better boost speed of the more recent CPU).
Grant
Darwin NT
ID: 2036607 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 2036619 - Posted: 7 Mar 2020, 23:55:49 UTC

You might try setting the BOINC Manager to 87.5 percent. That should allow your system to perform maximum production. When you crowd too many CPU tasks on your system you can slow it big time by running 99/100 percent CPU.

While it is true your CPU speed is lower than some. I suspect it can average below two hours per task rather than the 3+ hours that it appears to be running at now.

You will also be happier if you drop in at least a gtx 750ti on the system ($40USD on eBay).
Tom
A proud member of the OFA (Old Farts Association).
ID: 2036619 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13797
Credit: 208,696,464
RAC: 304
Australia
Message 2036623 - Posted: 7 Mar 2020, 23:58:41 UTC - in response to Message 2036619.  
Last modified: 7 Mar 2020, 23:59:59 UTC

You might try setting the BOINC Manager to 87.5 percent. That should allow your system to perform maximum production. When you crowd too many CPU tasks on your system you can slow it big time by running 99/100 percent CPU.
The improvements gained generally don't offset the loss in production from those unused threads.
Unless you are doing other CPU intensive things on a system that take CPU time from Seti, using all cores & threads will give the most output per hour.


You will also be happier if you drop in at least a gtx 750ti on the system ($40USD on eBay).
Unless you plan to move to another project that can make use of it in a few weeks, i don't see any benefit to adding a GPU to the system unless it costs nothing.
Grant
Darwin NT
ID: 2036623 · Report as offensive
6BQ5

Send message
Joined: 7 Dec 18
Posts: 29
Credit: 12,725,636
RAC: 357
United States
Message 2036693 - Posted: 8 Mar 2020, 5:41:08 UTC

Wow-wee, ka-za-wee! :)

6 to 8 weeks to really register the true performance of The-Beast? That's not going to happen now with the upcoming hibernation. I'm going to be installing two 2697 v2 processors in The-Beast for a bit more crunch power. :(

Yes, I have always noticed that GPU tasks on my C1-PC finish in under 6 minutes whereas CPU tasks could take 2 to 3 hours. Are the tasks by themselves for the CPU and GPU basically the same or are they completely different in nature? I always thought they were somehow different and that partially explained why the GPU tasks finish so quickly.

I am indeed hoping to apply my computers to other distributed projects but am not sure which ones specifically. So, a GPU is definitely in the plan for The-Beast. I was actually thinking of getting one of those new snazzy GeForce 20xx cards for the C1-PC and trickle the current GeForce 1080 card into The-Beast. If a lower end Quadro card is better then I could look into that option too. No matter what, when SETI comes back :) I'll be ready!

-=- Boris
ID: 2036693 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13797
Credit: 208,696,464
RAC: 304
Australia
Message 2036694 - Posted: 8 Mar 2020, 6:01:13 UTC - in response to Message 2036693.  

Yes, I have always noticed that GPU tasks on my C1-PC finish in under 6 minutes whereas CPU tasks could take 2 to 3 hours. Are the tasks by themselves for the CPU and GPU basically the same or are they completely different in nature?
There are only MB (MutiBeam) & AP (AstroPulse) Work Units. They "become" CPU or GPU when they are allocated by the server to the system doing the processing. If you look at your task list you will see WUs that you have processed with your GPU are Validated against ones processed by both GPUs & CPUs (and visa versa).
Grant
Darwin NT
ID: 2036694 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 2037385 - Posted: 11 Mar 2020, 23:08:32 UTC - in response to Message 2036623.  
Last modified: 11 Mar 2020, 23:09:52 UTC

You might try setting the BOINC Manager to 87.5 percent. That should allow your system to perform maximum production. When you crowd too many CPU tasks on your system you can slow it big time by running 99/100 percent CPU.
The improvements gained generally don't offset the loss in production from those unused threads.
Unless you are doing other CPU intensive things on a system that take CPU time from Seti, using all cores & threads will give the most output per hour.


When I was running a 40 thread Intel server if something hiccuped and the cpus started running 100% then the speed of the CPU task processing dropped significantly. Maybe I am missing something. If you are only running cpu tasks then run at 100%?

If you are running a mix of cpu and gpu tasks then the 90% and/or 2-4 threads left idle seems to work better.

The only hole in the argument that I can see is if you run 100% CPU threads on a CPU only system it will still slow down because the "other" tasks would be competing with Seti.

Oh, well. Since I don't have a CPU-only rig I can't confirm that empirically.

Tom
A proud member of the OFA (Old Farts Association).
ID: 2037385 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 2037524 - Posted: 12 Mar 2020, 11:42:57 UTC - in response to Message 2036603.  
Last modified: 12 Mar 2020, 11:46:03 UTC

Does the GPU card really make that much of a difference?

Thoughts?

-=- Boris

Yes, it does.

And second-order effects to consider: More simultaneous tasks mean increased CPU cache evictions/misses. And, finally,memory bandwidth saturation.
So, increasing number of logical CPUs w/o corresponding improvement in CPU L3 cache size and RAM speed will give diminishing results at some point.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 2037524 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 2037527 - Posted: 12 Mar 2020, 11:48:41 UTC - in response to Message 2036693.  

Are the tasks by themselves for the CPU and GPU basically the same or are they completely different in nature? I always thought they were somehow different and that partially explained why the GPU tasks finish so quickly.

-=- Boris

Exactly the same.
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 2037527 · Report as offensive

Message boards : Number crunching : More cores, worse performance? Why?


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.