I am having some power-related issues at my server hosting 2 Bluefields. Both of my Bluefield-2 DPUs are E-series, meaning they get power only through the PCIe slot. The server has two PSUs but since it is a NUMA architecture and also hosts an NVIDIA A100 :), when I connect the Bluefields with one cable, everything is fine. But when both ports are connected, I see
dmesg messages about the current drawn dropped to 27W or something like that.
Anyway, I was thinking that switching the mode of operation of one of the Bluefield back to SEPARATED_HOST mode might reduce the power consumption as the ARM cores do not need to handle packets.
- Would this be the case?
The other question i have is whether there is any tool to measure the power consumption of the Bluefield('s PCIe slot)?
Currently, I am using
ipmitool to see some general consumption stats, but they are not really reliable/accurte.