9th of November

USB charging
The Comsol USB charger that I wrote about in my last post turned out to be a dud. After about 24 hours I found that it had stopped one of the fans on top of the Pi^4 case and was trying to power off one of the Pi3B+ nodes. Clearly its designed for charging rather than providing constant power. I have put back the Astrotek USB charger which got all 4 Pis and the fans running. This is one of the drawbacks of using USB chargers, most of them aren’t designed to supply constant power.


An experiment
I thought I would try a couple of the Pis running just 3 tasks at a time instead of the usual 4 to see if it was any quicker. The short answer is yes. The longer answer is yes its quicker but not by enough.

After averaging out the run times the couple of Pis that ran 3 tasks came in about 15% quicker than those that ran 4 at a time. That isn’t enough to make up for the 4th core. The testing was on Pi3B+ nodes running the Einstein BRP4 work units which typically take 11 hours and 10 minutes (per task) when running 4 at a time. I haven't tried the Pi4 - It could give totally different results.


Aarch64
It seems the Raspberry Pi foundation have an experimental 64 bit kernel available that they’re testing, particularly on the Pi4 but the Pi3’s can also run it. There have been reports of issues getting it to run 32 bit apps (the armhf architecture) over at Einstein.

Ubuntu have also said they are going to support the Pi4 from Ubuntu 19.10 onwards. They also have noted a kernel bug where the USB stops working on the 4GB Pi4. Their work around is to tell the Pi it only has 3GB of memory until they can get the kernel patched.

26th of October

This last fortnight has been business as usual, except for the last two days.


The last two days
Its been hot so I set the cluster to NNT (no new tasks) and let it finish off what it had, which can take up to 20 hours for the slower Pis. I also set the cache levels to their minimum so it doesn't take so long to finish off the work on hand. Normally I keep 8 tasks per Pi, 4 in progress and 4 cached. The cluster has been idle since then apart from me taking the opportunity to apply updates. Its going to be hot again today.


Replacement USB charger
I got another 5 port USB charger rated at 39 watts/8 amps. Its a Comsol brand so lets see how long they last. That has freed up one of the Astrotek brand chargers which is now a spare. Given I had two of these Astrotek chargers fail I think it best to keep a spare.

I put the two powered-off Pi3B+ nodes on the new charger. They had been off since the last charger died. That puts Marks Rpi Cluster back to 12 x Pi3B+ and 10 x Pi3B compute nodes.

13th of October

Pi4 support nodes
The official power supply for the Pi4's finally arrived so I installed them this weekend.

The first problem I encountered was the cables for the PiDrive. The PiDrive has a Y split Micro USB cable for power and the Pi4 uses USB C. I had an 3TB external hard disk that has its own power adapter so that was one. For the other one I bought a portable SSD drive which came with a USB A to USB C cable and didn't require external power. 


First the proxy server
I then had to find my instructions for moving the root partition from the Micro SD card onto the external device. Most external drives come formatted as exFAT. My instructions uses the command line to do everything. For wiping the external drives I found its easier to plug them into one of my larger Linux machines and use gparted. Once I had them repartitioned and formatted I plugged them into the Pi.

I clean installed Raspbian. I probably could have just swapped the Micro SD cards from the old ones but instead I went with a clean install. I kept the old proxy server running so I could cut and paste the config file across. A couple more commands to set it up and its ready. I then had to go through each of the compute nodes and point all 20 of them to the new proxy server.

There were a bunch of Raspbian updates this weekend so the proxy server got a bit of a work out. File transfer speeds are higher than before for the cached content. I'm not sure if that is the result of the faster network port or using the SSD. I don't think its the extra memory as the proxy doesn't seem to be using it at the moment.


Next the NFS server
The second support node is a NFS server. Again I had to dig out instructions on how to set it up. I did the same process to move the root partition but this time I used the 3TB external hard disk. I kept the old NFS server running so I could copy the files across before powering the old one down for the final time. In hind sight I probably should have done the NFS server first and then I could have just copied the configuration files rather than editing them on each of the compute nodes.

For cooling I left the top off the case. I need to look at a better solution. Unlike the compute nodes the CPU isn't working hard so they aren't running hot.


Current cluster
Marks Rpi Cluster now consists of:
12 x Pi3B+ as compute nodes (two not working due to a failed USB charger)
10 x Pi3B as compute nodes
2 x Pi4B as support nodes

Progress as of 7th of October

Current status
Marks Rpi Cluster continues to run 24/7. The cluster currently has 10 Pi3B+ and 10 Pi3B's all doing Einstein BRP4 work.

The cluster is up to 6.8M credits for Einstein with a recent average of 9,673 credits per day.


Other news
There are two more Pi3B+ nodes off-line which I haven't included in the above numbers because their power supply (a USB charger) failed.

I am still waiting for the back ordered USB-C power supplies for the Pi4's which will become support nodes.