rkanagar
commented
1 month ago Hi wangerforcs, The perf_monitoring.c has all the code of monitoring by perf. It uses standard system call perf_event_open which is defined in lib/perf.c
https://www.systutorials.com/docs/linux/man/2-perf_event_open/ https://github.com/intel/intel-cmt-cat/blob/master/lib/perf_monitoring.c
Events:
/** List of non virtual events */
const enum pqos_mon_event mon_event[] = {
PQOS_MON_EVENT_L3_OCCUP,
PQOS_MON_EVENT_LMEM_BW,
PQOS_MON_EVENT_TMEM_BW,
PQOS_PERF_EVENT_LLC_MISS,
PQOS_PERF_EVENT_LLC_REF,
(enum pqos_mon_event)PQOS_PERF_EVENT_CYCLES,
(enum pqos_mon_event)PQOS_PERF_EVENT_INSTRUCTIONS,
PQOS_PERF_EVENT_LLC_MISS_PCIE_READ,
PQOS_PERF_EVENT_LLC_MISS_PCIE_WRITE,
PQOS_PERF_EVENT_LLC_REF_PCIE_READ,
PQOS_PERF_EVENT_LLC_REF_PCIE_WRITE};In the above events, the required event is monitored every second.
wangerforcs
My processor is an Intel(R) Xeon(R) Platinum 8269CY CPU with an L3 cache size of 35MB. I've noticed that this metric LLC occupancy is a multiple of 104KB, and in many cases, there's a significant discrepancy between the size calculated based on the cache ways allocated by CAT (Cache Allocation Technology) and the actual data monitored. Therefore, I would like to gain a deeper understanding of how this data is derived.