Add resctrl collector

[rtreffer]

Expose the Linux resctrl monitoring counters (Intel RDT, AMD PQoS, ARM MPAM): per L3 domain memory bandwidth (mbm_total_bytes, mbm_local_bytes) and last level cache occupancy (llc_occupancy). Parsing is done by the procfs resctrlfs package.

The collector is disabled by default because the resctrl filesystem must be mounted first (mount -t resctrl resctrl /sys/fs/resctrl). The mountpoint is configurable via --collector.resctrl.path.

Counters a CPU does not support (for example mbm_local_bytes on Arm MPAM) or that read "Unavailable" are skipped per sample. The hardware counters wrap; Prometheus treats the wrap as a counter reset, so no correction is applied.

Depends on a procfs release that contains the resctrlfs package (prometheus/procfs#862).

Signed-off-by: René Treffer treffer@measite.de

Reviewers

AI disclosure: This code was written with the help of AI.
I have read through the code and verified that it works

Here is a snippet showing the metrics as exported by node_exporter built against prometheus/procfs#862

# HELP node_resctrl_llc_occupancy_bytes Last level cache bytes occupied in this domain.
# TYPE node_resctrl_llc_occupancy_bytes gauge
node_resctrl_llc_occupancy_bytes{domain="0"} 3.2014336e+07
node_resctrl_llc_occupancy_bytes{domain="1"} 2.3003136e+07
# HELP node_resctrl_memory_bandwidth_bytes_total Bytes moved between the last level cache and memory. scope=total includes remote sockets, scope=local is memory attached to this domain.
# TYPE node_resctrl_memory_bandwidth_bytes_total counter
node_resctrl_memory_bandwidth_bytes_total{domain="0",scope="local"} 9.866260884132983e+18
node_resctrl_memory_bandwidth_bytes_total{domain="0",scope="total"} 1.8399210738624e+13
node_resctrl_memory_bandwidth_bytes_total{domain="1",scope="local"} 5.010254585456994e+18
node_resctrl_memory_bandwidth_bytes_total{domain="1",scope="total"} 4.003923930496e+12

You can use sum by (instance,node) (rate(node_resctrl_memory_bandwidth_bytes_total{scope="total"}[1m])) to get an estimate of the data transferred between memory and CPU.

This is particularly interesting for workloads like CPU inference, in-memory databases or vector databases.

Caveats: The total value is only available if memory access happened. The metric won't be available on every scrape. We could cache the latest value if this is an issue.