Saturday, November 5, 2022

5th of November

Marks Rpi Cluster continues to run 24/7. After my last post we had a few hot days so I allowed the cluster to finished off its current work and idle. It now back to crunching again although today was hot as well with the room they are in reaching 30 degrees C.

With the break in crunching for a few days my RAC (Recent average credit) has dropped from from 40k to 29k. The cluster is up to 19.7M credits so at the current rate it might take 3 days to hit 20M credits, assuming I don't stop again due to the heat.

Seeing as I can't get a hold of any more Raspberry Pis I have been looking at the Raxda Rock 5B online. It is an ARM single board computer, almost double the size of the Rpi and comes in a Pico-ITX format. It has 4 Cortex A55 cores, 4 Cortex A76 cores, a Mali G610 GPU and ton of connectivity. Given its combination of cores its probably best to crunch on the A76 cores and leave the A55 cores to run less demanding work. They come in 4, 8 and 16GB memory sizes. Unfortunately it doesn't have the software support that the Pi has so a number of things don't work under Debian. It can also run Android.


Sunday, October 23, 2022

23rd of October

Marks Rpi Cluster continues to run 24/7. The cluster has 19.4M credits accumulated for Einstein@home and has a RAC (Recent Average Credit) of 40k. If output continues at the current pace it should hit 20M credits in a bit over a fortnight.

I would like to make the cluster a bit more resilient by duplicating a couple of the support nodes. However I can't seem to find the Pi4 2GB model for sale unless I resort to the overpriced ones on eBay. They make approximately 500k units a month according to Eben Upton but they still seem impossible to find. Their advice was to keep checking rpilocator.com to see where stock is available.


Sunday, October 9, 2022

9th of October

Marks Rpi Cluster continues to run 24/7. As mentioned in previous posts its concentrating on running Einstein@home BRP4 work. That is 24 Raspberry Pi4's looking through data captured from the Arecibo radio telescope (before its demise), searching for Binary Radio Pulsars.

Most of the compute nodes have managed to get their RAC (Recent Average Credit) up to 1600, with a couple even getting up to 1700.

I was asked last week what the up-time on the compute nodes was. These days it is usually one to two weeks due to security fixes and other updates, but I have gone for a few months before rebooting them.


Universe@home
I had a brief look at running Universe@home on Marks Rpi Cluster. They are an astronomy project currently doing analysis on Black Holes.

Their Raspberry Pi app is for the ARMv6 (original Raspberry Pi). The Pi2 was ARMv7 and the Pi3 and Pi4 are both ARMv8. In other words the app is not able to take advantage of the newer Pis hardware. Its also a 32 bit app so won't take advantage of a 64 bit operating system and one has to install support libraries to run them under Raspberry Pi OS 64 bit.

Needless to say I gave up on the idea of running it.

Saturday, September 3, 2022

Progress as of 3rd of September

This is where Marks Rpi Cluster is as of the 3rd of September for Einstein@home. Total credit is 17,387,442 and recent average is 31,909 credits.

 

Einstein@home had a power outage earlier in the week and apparently the Atlas Cluster that they use had UPS failures and they haven't managed to revive enough of the cluster. Its used to pre-process the BRP4 work so they have run out. In addition Rosetta@home doesn't have any work, but they haven't for months.

The dip on the graph just before the word Today is when they they went down earlier in the week and the spike hitting the word Today was Marks Rpi Cluster returning completed work after the initial outage. Normally I keep 0.3 days of work cached, but given the recent instability of the project I will be increasing it.


Cluster make up
As of the 3rd of September the cluster is made up of:

3 x Pi4 (2GB) as support nodes
24 x Pi4 (8GB) as compute nodes

For those that like numbers that is 96 ARMv8 CPU cores, 192GB of memory and 768GB of SD card storage for the compute nodes.

The support nodes are a proxy server, a storage server and a time server with a GPS.