VMW Research Group Raspberry Pi Cluster Mark2
Summary
We have created a 5-node Raspberry Pi 4 cluster,
with 20 cores and 24GB of RAM.
This replaces our earlier Pi-2 cluster
that had 96 cores with 24GB of RAM (the new cluster draws less power
while being 3x faster).
The cluster draws around 65W when under load, and around 33W when idle.
We use SLURM for job submission.
It obtains 50.3 GFlops of performance while running HPL Linpack,
at 0.86 GFLOPS/W.
================================================================================
T/V N NB P Q Time Gflops
--------------------------------------------------------------------------------
WR02R2L2 40000 256 5 4 841.08 5.0732e+01
Plot of power usage while running Linpack:
Publications
TODO
- Better visualization, add more LED screens
- Expand to 8 nodes? Would need an additional router
- Get libpfm4/PAPI support going for Cortex A72 / Pi4
Pictures
Watts Up Pro showing power usage:
Views of the new cluster:
Old cluster and new cluster together:
Documentation
Full cluster build instructions, code used, etc, will be posted RSN.
Cluster Construction Progress
12 March 2024
With spring break here
finally got slurm and Linpack set up and running. Reliably comes up
after reboot, which is nice (the old cluster didn't do that consistently).
1 January 2024
Was thinking of upgrading the old Pi cluster to have Pi4s. Had been putting
off upgrading because no matter when I start buying Pis to do it the
Pi Foundation would immediately release a new board (Pi1B, Pi2B, Pi2B+,
Pi3, Pi4, etc). They did announce the Pi5 but I don't think those would
be suitable for a cluster as this point.
The big issue with the old one was Linux was way out of date (had trouble
migrating accounts for the ECE574 Cluster Computing students as the new
password hashes too new for the version of Debian running). Upgrading in
place difficult too as they only had 4GB SD cards.
Now that the parts shortage finally let up decided to go with Pi4s and
Power over Ethernet for the power.
Got Linux installed and the hardware all going, but ran out of time to
complete things before the semester started.
See the original Pi Cluster.
Back to projects page.