Current status
Marks Rpi Cluster continues to run 24/7. The cluster currently has 12 Pi3B+ and 4 Pi3B's all doing Einstein BRP4 work.
The cluster is up to 6.09M credits for Einstein with a recent average of 7,735 credits per day.
Raspbian Buster rollout
As I mentioned in my previous post I have started updating the cluster from Raspbian Stretch to Buster (the current release). I have done 12 out of 16 compute nodes. The last four are in a difficult access location. I have also done 1 of the 2 support nodes.
Pi4 update
The two Pi4 model B's are still on back-order with an ETA of September. These will become the support nodes. On the same order are two USB C power supplies and two official cases. I have also ordered Noctua 40mm fans which have a higher airflow than the current ones.
3 comments:
Hi Mark,
I stumbled on your blog in my search of Raspberry Pi cluster knowledge (specifically relating to SETI@Home). Your setup is impressive - do you have a master "behind the scenes" from your clusters or is it one of the nodes? Also, I continue to read your posts but have not seen any mention of software. Are you using some version of mpich? I know BOINC works a bit differently since it is already designed for distributed computing, but do you also use a scheduler for BOINC? What about the GPU? I have read about VC4CL but not much about using it for distributed computing.
Any advice would be greatly appreciated. I have found several detailed guides but nothing as thorough as your blog.
Lets be clear, they are just a bunch of Raspberry Pis, sometimes called a Cluster of Workstations or CoW. There is no controlling node, each Pi is running BOINC and doing its own thing, they don’t interact at all. They’re all headless, that is they don’t have a screen or keyboard plugged in. Each compute node Pi has standard Raspbian lite and the BOINC client installed, nothing else needed.
I’ve mentioned in a few posts that I have a couple of support nodes. One is a network proxy so it handles all the internet traffic. The second one is a storage server where I put things like config files and the like that I use when installing or reinstalling the Pi’s. Both of the support nodes have a Pi Drive attached.
I control the compute nodes using BOINCtasks on a Windows machine. I can see what each Pi has in the way of work as well as seeing all of them together. It can do all the usual things you do via BOINC Manager like suspending tasks, no new work, etc.
As for VC4 it never took off. The Pi was too limited in memory and what it could do so we don’t use it. Maybe the VC6 that the Pi4 has will be usable in time.
Thank you! This is exactly the kind of information I was searching for. I figured BOINC could already do a better job at distributing work than anything I could come up with. Still, I am curious about somehow harnessing the power of the GPUs.
Post a Comment