My ARM fleet

[Dirk Broer]

Do you have one or more ARM SBC's and do you have problems selecting the right BOINC project?
Schermafbeelding 2023-06-28 012539.jpg
A word of warning: having ARM Applications on a project does not mean that you'll be able to run them on your board, -unless it is WEP-M+2.
It looks like some developers test it on systems with either more RAM than 4GB, or only test it with one WU running and completing and then declare it OK for the wild.

Cases:
Drowning your Odroid-N2+ (six cores, 4GB on my case) in the various 64-bit ARM apps leads to a complete standstill in a week.
Same for a 4GB Rock Pi 4B (six cores, 4GB).
Same for two 4GB Raspberry Pi 4's (four cores, 4GB).
Same for a 4GB Jetson Nano (four cores, 4GB).
I didn't even try with my 2GB Jetson Nano....let alone my 1GB Raspberry Pi 3's.

Stupid thing is that the 32-bit control group (an octa-core 2GB Odroid XU4, so a mere 256MB per core) has no problems at all.

[Dirk Broer]

Stupid thing is that the 32-bit control group (an octa-core 2GB Odroid XU4, so a mere 256MB per core) has no problems at all running the 32-bit ARM/Linux apps.

That finding was the start for "The return of the 32-bit squadron".

As the Odroid-XU4 effortlessly crunched its way through Universe@Home, SiDock@Home and Asteroids@Home, running all these three projects along each other without any problems I paid closer attention to it. And while Asteroids@Home -the very same project that crashed half the 64-bit squadron- paid s*#t, SiDock@Home paid better and Universe@Home appeared downright generous with credits!

So the experimental department quickly refurbished the ASUS Tinker Board, to crunch only Universe@Home. And where WEP-M+2 in the past paid a daily 900-1,000 credits, Universe@Home now does around 9,000!

I then stopped Asteroids@Home for the XU4 and I decided to re-launch my 64-bit Cortex-A53 Raspberry Pi 3B's for Universe@Home and for SiDock@Home. The Universe tasks soon had the same success, credits per Watt-wise. The SiDock tasks had less success -perhaps the 1GB of RAM was not really enough, or the Cortex-A53 not strong enough, who knows? So I stopped SiDock for the Pi 3B's and converted them to do only Universe as well.

Based upon their scores I converted my 64-bit Cortex-A55 Odroid-M1 to be able to crunch 32-bit Universe@Home apps too -you never see Universe 64-bit Aarch64 tasks-, and it is now on its way to best the Tinker Board in its daily scores!

Now my 32-bit Cortex-A15 BeagleBone-AI is at it too, and the 64-bit Cortex-A72/A53 Rock Pi 4B is waiting in the docks to be refitted, as are the 64-bit Cortex-A57 nVidia Jetson Nano's. I might even want to re-launch the 32-bit Cortex-A9 Hummingboard and the 32-bit Cortex-A5 Odroid-C1....at least temporary. If that factor nine times more credit holds also true for 64-bit boards running 3-bit tasks, then the Radxa Rock Pi 4B must be able to reach 27,000 credits a day -and the Raspberry 4's more than 30,000 per day...

[vaughan]

So Universe is the best for credits with SBCs?

[Dirk Broer]

So Universe is the best for credits with SBCs?

It very well looks like it. As per request I 'stopped' WEP-M+2, and went searching for a (more) profitable project. It was really just by chance I did try my ASUS Tinkerboard on Universe@Home and was rewarded with a nine-fold score as compared to WEP-M+2!
After that I recommissioned my Odroid-XU4 -as mentioned earlier- and bought a BeagleBone-AI. As they performed very well on Universe@Home too, the Raspberry Pi 3B's and the Odroid-M1 were converted as well and just now I just have my Jetson Nano's pointed to Universe.
Next will be my Rock Pi 4B, after that the Raspberry Pi 4 fleet, my Odroid-N2+ and my Jetson Xavier-and I just have bought a Pi 5....

In theory Universe@Home has 64-bit Aarch64 work too, but in practice you only encounter the 32-bit ARMv7 (and ARMv6!) BHspin v2 app.

[Dirk Broer]

For those of us that have 64-bit Linux/ARM systems and that want to crunch 32-bit apps here as reminder

Your 64-bit OS has standard no 32-bit libraries! ARMv8.0-A and ARMv8.2-A systems -and quite likely any other ARMv8.x-A- can be 'tricked' into accepting 32-bit work though by this command-line 'hacking' (debian-style):

sudo dpkg --add-architecture armhf
sudo apt update --fix-missing
sudo apt dist-upgrade
sudo apt install libc6:armhf libstdc++6:armhf zlib1g:armhf libfuse2:armhf libgomp1:armhf libboinc7:armhf

to be sure you get all, edit your cc_config.xml to have these lines on a ARMv8 system:

<cc_config>
    <options>
        <alt_platform>arm-unknown-linux-gnueabihf</alt_platform>
        <alt_platform>armv7l-unknown-linux-gnueabihf</alt_platform>
    </options>
</cc_config>

Mind you: with the newest and latest upcoming ARMv9 SOCs running 32-bit taks will no longer be possible! (only applies when you are crunching on the very latest high-end smartphones: Cortex-A510, Cortex-A710, Cortex-A715, Cortex-X2, Cortex-X3, Neoverse N2, Neoverse V2)

If you run a 64-bit OS on a Raspberry Pi 3, or a 4 with 1 or 2 GB of RAM (perhaps even with 4GB), be sure to install ZRAM as well:
add the following lines to /etc/sysctl.conf (and then enable in real-time with the appropriate commands.)

vm.vfs_cache_pressure=500
vm.swappiness=100
vm.dirty_background_ratio=1
vm.dirty_ratio=50

To enable these settings temporarily without rebooting, use the following commands:

sudo sysctl -w vm.swappiness=100
sudo sysctl -w vm.vfs_cache_pressure=500

[Dirk Broer]

How's the 32-bit squad doing on Universe@Home, and how do their results compare to the converted 64-bit 'heavies'?

• BeagleBone-AI (dual core Cortex-A15): 4,035 credits/day
• Asus Tinker Board (quad core Cortex-A17): 8,086 credits/day
• Odroid-XU4 (quad core Cortex-A15/quad core Cortex-A7): 8,923 credits/day

These values hover a bit for the Tinker Board and the XU4 (between 7800 and 9100), but the BeagleBone-AI is still climbing slowly.
I quitted crunching Asteroids@Home on the Odroid-XU4, so it only crunches SiDock@Home and Universe@Home now.

Longer crunching 64-bitters, still on the rise and climbing a tad faster than the BeagleBone-AI (especially the Odroid-M1)

• Raspberry Pi 3 (quad core Cortex-A53): 2,772 credits/day
• Raspberry Pi three (quad core Cortex-A53): 4,421 credits/day
• Odroid-M1 (quad core Cortex-A55): 9,354 credits/day

The Raspberry Pi 3 previously crunched SiDock@Home but the Odroid-XU4 is much better in that, so the Pi 3 went to Universe@Home.

and here for the last converted 64-bitters, that began later with Universe@Home and climb still faster:

• Raspberry Pi 400 (quad core Cortex-A72): 3,409 credits/day
• Raspberry Pi 4 8GB (quad core Cortex-A72): 3,480 credits/day
• Jetson Nano 2GB (quad core Cortex-A57): 4,059 credits/day
• Jetson Nano 4GB (quad core Cortex-A57): 6,680 credits/day

The 2GB Jetson Nano was the lastly converted and much celebrated at the experimental department, as it now combines -thanks to Muhammad Yunus,
IoT Engineer, Software Developer & Machine Learning Enthusiast- both CUDA and OpenCL for its GPU.
For the time being it is as useful as a 100 TOPS NPU/TPU for BOINC running Linux/ARM (or even Windows), but who knows....
The two Raspberry Pi 4's are the latest additions to the Universe task force

[Dirk Broer]

As SiDock@Home is presently without work the Odroid-XU4 now runs Universe@Home exclusively, so the credit tables have been shaken up.
I also keep converting 64-bit ARM SBCs to take 32-bit work, so the list is getting longer and longer and values change daily

The top-4 ARM systems, pure credit-wise:

1. Odroid-XU4 13,019
2. Raspberry Pi 4 8GB 12,310
3. Raspberry Pi 400 11,037
4. Odroid-M1 9,597

The top-4 ARM systems, credit per core wise:

1. Raspberry Pi 4 8GB 3,078
2. Raspberry Pi 400 2,759
3. Odroid-M1 2,399
4. BeagleBone-AI 2,268

The top-4 ARM systems, credit per Watt wise:

1. Raspberry Pi 400 1,698
2. Odroid-M1 1,391
3. Raspberry Pi three 1,227
4. Raspberry Pi 4 8GB 1,001

The top-4 ARM systems, credit per core per Watt wise:

1. BeagleBone-AI 454
2. Raspberry Pi 400 425
3. Odroid-M1 348
4. Raspberry Pi three 307

[Dirk Broer]

That finding was the start for "The return of the 32-bit squadron".

As the Odroid-XU4 effortlessly crunched its way through Universe@Home, SiDock@Home and Asteroids@Home, running all these three projects along each other without any problems I paid closer attention to it. And while Asteroids@Home -the very same project that crashed half the 64-bit squadron- paid s*#t, SiDock@Home paid better and Universe@Home appeared downright generous with credits!

So the experimental department quickly refurbished the ASUS Tinker Board, to crunch only Universe@Home. And where WEP-M+2 in the past paid a daily 900-1,000 credits, Universe@Home now does around 9,000!

I then stopped Asteroids@Home for the XU4 and I decided to re-launch my 64-bit Cortex-A53 Raspberry Pi 3B's for Universe@Home and for SiDock@Home. The Universe tasks soon had the same success, credits per Watt-wise. The SiDock tasks had less success -perhaps the 1GB of RAM was not really enough, or the Cortex-A53 not strong enough, who knows? So I stopped SiDock for the Pi 3B's and converted them to do only Universe as well.

Based upon their scores I converted my 64-bit Cortex-A55 Odroid-M1 to be able to crunch 32-bit Universe@Home apps too -you never see Universe 64-bit Aarch64 tasks-, and it is now on its way to best the Tinker Board in its daily scores!

Now my 32-bit Cortex-A15 BeagleBone-AI is at it too, and the 64-bit Cortex-A72/A53 Rock Pi 4B is waiting in the docks to be refitted, as are the 64-bit Cortex-A57 nVidia Jetson Nano's. I might even want to re-launch the 32-bit Cortex-A9 Hummingboard and the 32-bit Cortex-A5 Odroid-C1....at least temporary. If that factor nine times more credit holds also true for 64-bit boards running 3-bit tasks, then the Radxa Rock Pi 4B must be able to reach 27,000 credits a day -and the Raspberry 4's more than 30,000 per day...

Looks like the ninefold credit only holds true when you compare ARM 32-bit system WEP-M+2 scores to ARM 32-bit system Universe scores.
For ARM 64-bit (Aarch64) it is 4 to 4.5 times. I found out that my Raspberry Pi 3's levelled out at between 3,000 and 4,500 credits. They would fetch 500-550 credits at WEP-M+2 when on 32-bit Raspbian while they scored 1,000-1,100 running 64-bit Raspberry Pi OS.
The Odroid-M1 that always ran 64-bit Debian scored around 2,000 to 2,200 credits on WEP-M+2, while it now hovers between 9,500 and 10,000 credits for Universe.

64-bit systems also seem to have benefits of more RAM, while the 32-bit Odroid-XU4 keeps lustily hammering away at Universe@HOMe on a mere 2GB for its eight cores.
At the 64-bit systems the 8GB Odroid-M1 outperforms the 4GB Jetson Nano, while that outperforms the 2GB Jetson Nano in its turn.
The 1GB Raspberry Pi 3's are the lowest 64-bit performers of the field -but have just as much RAM per core as the stellar performing 32-bit Odroid-XU4....

[Dirk Broer]

With both WEP-M+2 and Universe@Home down and/or out the fleet has turned its attention to SiDock@Home.

SiDock is far tougher though on the SBCs than WEP-M+2 or Universe@Home. Like Universe@Home it has presently only 32-bit work, but it has more needed requisites.

The Odroid-C1, Hummingboard-i4ex and both Raspberry Pi 3Bs bit the dust on SiDock and are waiting for the revival of Universe@Home.

My Jetson Nano 2GB gets no work done on account of its L4T Linux version is too old (nVidia standard adaption of Ubuntu 18.04.6 LTS with kernel 4.9.201) and only has GLIBC 2.27, while 2.28 or higher is needed.

My Jetson Nano 4GB gets no work done on account of missing libgomp1:armhf and my Jetson Xavier NX and my Raspberry Pi CM4 ran out of diskspace with SiDock.

The rest off the fleet does well though, it is not the debacle that I faced trying to run the 64-bit Asteroids@Home application on all cores of the SBCs.

[vaughan]

PrimeGrid is supposed to have Linux ARM64 applications for GFN16 and GFN17 but I cannot get them to get work on my Odroid.

Only works for Asteroids and SiDock (and the NCI WuProp).