News:

Members can see all forum boards and posts. Non members can only see a few boards.  If you have forgotten your password use this link to change it.
https://forum.boinc-australia.net/index.php?action=reminder

Main Menu

How do you we measure performance?

Started by Lennox84, June 05, 2023, 10:03:02 AM

Previous topic - Next topic

Lennox84

Was hoping to tap the brain trust here and understand how do you read the results in Boinc Manager, or what tools do you use (preferably in Linux) to identify how the system is performing??

I think I've seen the term hash rate thrown around, but I'm not sure if that's here or in unrelated documents.

I have a few machines in various generations Intel Gen 3,4,6 & AM4 also to mobile phones just for boinc (24/7).

I'd love to better understand the difference between them in a performance metric to understand what to look out for on when I perform an upgrade to the system or when the next "Facebook Marketplace Deal" comes along to I need a good excuse to maximise how I contribute and try to keep the hobby economical-ish (In case the wife ever asks)...

Would love to hear anyone's tips & tricks in relation to setting up a BOINC system, what processors are better for what etc...



MiniFarm:
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G3 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)

Laptops:
Yoga C940 - i7-1065G7 - 16GB - NVME - Iris iGPU +  GTX 680 (Windows)
FrameWork - i5-1240p - 32GB - NVME - Iris XE iGPU (Fedora + Docker) Part-time Cruncher

Server:
B550 Aorus Elite - AMD 3700x - 32GB - Quadro k4000- (UNRaid - Docker)

Desktop:
X570 I AORUS Pro WiFi - Ryzen 3600 - 32GB - RX 5700XT - (Fedora + Docker)

Offline -
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
Lenovo m700 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - SSD - No GPU (Ubuntu Server)

Dark Angel

BOINC uses two benchmarks to rate a system's performance.  A Whetstone and a Dhrystone
.  tldr; it's a floating point test and an integer test.  It them uses these as well as the time taken to process workunits to derive a score that's used to determine how many "cobblestones" it will claim.  Whether the project grants that or gives credit on a system of it's own is entirely up to them.
In days of yore I used to keep a spreadsheet which tabulated various CPUs along with their installed operating systems, clock speed (stock of overclocked), the BOINC client version, and their scores.  It's long out of date now as I haven't kept it up in years.

chooka03

The only tip I can offer is that some projects run better with Linux than Windows. Universe@Home, LHC, Cosmology, all run better with Linux. Example, I run Linux on a virtual machine for Cosmology and it does the work units in half the time of Windows. There's that much difference.

All about cores really. Newer pc's with lots of core will perform the most work.The 7950X now uses AVX-512 which is a real bonus for Primegrid. Old Intel CPU's have used it for a while. AVX-512 is great for Primegrid.

As for GPU's, older gpu's with good FP64 are good for projects like Milkyway@Home. (Radeon VII's, Titan V's etc) FP32 is good for projects like Einstein.

I just measure the performance of my systems but running them say CPU only and letting it go for a few days. You'll soon see what each pc can produce.

This hobby used to be kinda economical but today's electricity prices have killed that (and high gpu prices)
New tech has proven to be more efficient though. My older cpu's pull around 200W vs newer one's around 150W (give or take)
My 4070ti is WAY more efficient than my 3070ti for example. (currently doing some Einstein and the 3070 pulls 280W vs the 4070 pulling 165W for example)

Lennox84

Quote from: Dark Angel on June 05, 2023, 10:45:22 AMtldr; it's a floating-point test and an integer test.  It them uses these as well as the time taken to process workunits to derive a score that's used to determine how many "cobblestones" it will claim.

So does this benchmark application look like it could be used? CoreMark
It does both Floating and Integer tests. It's command line only but the documentation seems easy enough to understand (Hopefully I can wrap my head around it.)


MiniFarm:
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G3 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)

Laptops:
Yoga C940 - i7-1065G7 - 16GB - NVME - Iris iGPU +  GTX 680 (Windows)
FrameWork - i5-1240p - 32GB - NVME - Iris XE iGPU (Fedora + Docker) Part-time Cruncher

Server:
B550 Aorus Elite - AMD 3700x - 32GB - Quadro k4000- (UNRaid - Docker)

Desktop:
X570 I AORUS Pro WiFi - Ryzen 3600 - 32GB - RX 5700XT - (Fedora + Docker)

Offline -
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
Lenovo m700 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - SSD - No GPU (Ubuntu Server)

Lennox84

Quote from: chooka03 on June 05, 2023, 05:08:38 PMUniverse@Home, LHC, Cosmology, all run better with Linux. Example, I run Linux on a virtual machine for Cosmology and it does the work units in half the time of Windows. There's that much difference.

All about cores really. Newer pc's with lots of core will perform the most work. The 7950X now uses AVX-512 which is a real bonus for Primegrid. Old Intel CPU's have used it for a while. AVX-512 is great for Primegrid.

I just measure the performance of my systems but running them say CPU only and letting it go for a few days. You'll soon see what each pc can produce.

This is also within my limited understanding and all my systems will be Linux based. Still figuring out wether I have a few Alpine Linux installs and run boinc in a container or run a few Ubuntu installs headless with boinc installed directly.

I'm running a few 6th gen HP 400's so I thought it would be a good time to try some remote management and boinc in a homelab kinda thing.


MiniFarm:
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G3 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)

Laptops:
Yoga C940 - i7-1065G7 - 16GB - NVME - Iris iGPU +  GTX 680 (Windows)
FrameWork - i5-1240p - 32GB - NVME - Iris XE iGPU (Fedora + Docker) Part-time Cruncher

Server:
B550 Aorus Elite - AMD 3700x - 32GB - Quadro k4000- (UNRaid - Docker)

Desktop:
X570 I AORUS Pro WiFi - Ryzen 3600 - 32GB - RX 5700XT - (Fedora + Docker)

Offline -
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
Lenovo m700 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - SSD - No GPU (Ubuntu Server)

Dark Angel

Quote from: Lennox84 on June 05, 2023, 07:16:20 PM
Quote from: chooka03 on June 05, 2023, 05:08:38 PMUniverse@Home, LHC, Cosmology, all run better with Linux. Example, I run Linux on a virtual machine for Cosmology and it does the work units in half the time of Windows. There's that much difference.

All about cores really. Newer pc's with lots of core will perform the most work. The 7950X now uses AVX-512 which is a real bonus for Primegrid. Old Intel CPU's have used it for a while. AVX-512 is great for Primegrid.

I just measure the performance of my systems but running them say CPU only and letting it go for a few days. You'll soon see what each pc can produce.

This is also within my limited understanding and all my systems will be Linux based. Still figuring out wether I have a few Alpine Linux installs and run boinc in a container or run a few Ubuntu installs headless with boinc installed directly.

I'm running a few 6th gen HP 400's so I thought it would be a good time to try some remote management and boinc in a homelab kinda thing.

I personally like my headless machines to run Ubuntu Server (recent version) with only the bare essentials installed, including no desktop.  I administer them via ssh ( cssh actually, saves a lot of typing when running the same commands across multiple machines) and run a local caching proxy (Squid) on an older i5 rig to save network traffic when I run LHC Native work as well as for caching update packages.
My headless rigs run Boinc from the Ubuntu repositories but my desktop has a locally compiled version.  I used to use the generic installer when I ran Slackware but that was a long time ago.
I get the impression from doing some reading that you'll get better performance running Boinc in Ubuntu than in Alpine.

Lennox84

So it's been a back and forth journey but I seem to be making some headway. A milestone I was personally shooting for was achieving 1,000,000 credits a month output which I've just recently achieved. I still need to do my AMD swap for my home server which has been offline with boinc for a week now ready for my Unraid upgrade (All tasks completed for a clean break & restart with the new hardware)

My mini farm has now evolved to 1x HP 400 G3, 1x HP 800 G2, 3 x Thinkcentre M710 All running i5 6500 or 6500T CPU's. & 2 x Mobile phones (Android 10)

Been adding my Framework 1240p and my AMD 3600 desktop (This will get a 5800x in the next week or so, I bought it but haven't installed it yet)

Ubuntu Server has been great to set things up Thanks @DarkAngel for the advice and feedback. Dingo gave me the idea to create a script too which makes things a whole lot easier.

I've been recording the Boinc Benchmarks to see how each system is working but I really need to get my GPU's working as my AMD desktop has a 5700XT which isn't being utilised even tho it's detected by the Boinc Flatpak and I have a K4000 in an EGPU enclosure which my windows install detects but isn't using  :banghead

Either way I'm slowly moving up the ranks and am very happy about it... but I feel like there's a fair bit more I need to do until I get the best results possible.


MiniFarm:
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G3 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)

Laptops:
Yoga C940 - i7-1065G7 - 16GB - NVME - Iris iGPU +  GTX 680 (Windows)
FrameWork - i5-1240p - 32GB - NVME - Iris XE iGPU (Fedora + Docker) Part-time Cruncher

Server:
B550 Aorus Elite - AMD 3700x - 32GB - Quadro k4000- (UNRaid - Docker)

Desktop:
X570 I AORUS Pro WiFi - Ryzen 3600 - 32GB - RX 5700XT - (Fedora + Docker)

Offline -
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
HP Elitedesk G2 - 6500 - 16GB - HDD - No GPU (Ubuntu Server)
Lenovo m700 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - NVME - No GPU (Ubuntu Server)
Lenovo m701 6500T - 16GB - SSD - No GPU (Ubuntu Server)

ryzenmulti

re: raw performance, less is more. not all linux flavours are fast, but the ones I've tested are faster than windoze. basic graphic user interface versions running xfce are quick. I use manjaro - it is 15% faster than windoze for the same projects. running VM's requires a dedicated cpu core to avoid lag. swap files are slow - the slowest part of any computation process is swapping in and out of memory, so combining minimal loaded o/s, good memory and decent cpu/gpu combination is the key IMHO. performance benchmarks return good results when the computer is doing nothing else. I have recently started leaving 1 full cpu core spare in Boinc manager just to manage other tasks/scheduling/file transfers and it seems to work better.

my 2c
The further back you look, the further forward you can see.

COMING SOON!! (2024)
136 intel cores (no H/T), 212 AMD/ryzen cores 8 RTX GPU's, 10 Tesla GPUs and 1.5TB RAM

Home cooked twin primes using python ... it started out with 256 digits of pi and eulers number ... and has ended with
(6*(3358638*(5^6137)+177))-1 ,4297 digits, is prime
(6*(3358638*(5^6137)+177))+1 ,4297 digits, is prime