cc-metric-collector/collectors/likwid/groups/haswellEP/FALSE_SHARE.txt
2021-03-25 14:47:10 +01:00

35 lines
1.4 KiB
Plaintext

SHORT False sharing
EVENTSET
FIXC0 INSTR_RETIRED_ANY
FIXC1 CPU_CLK_UNHALTED_CORE
FIXC2 CPU_CLK_UNHALTED_REF
PMC0 MEM_LOAD_UOPS_L3_HIT_RETIRED_XSNP_HITM
PMC1 MEM_LOAD_UOPS_L3_MISS_RETIRED_REMOTE_HITM
PMC2 MEM_LOAD_UOPS_RETIRED_ALL_ALL
METRICS
Runtime (RDTSC) [s] time
Runtime unhalted [s] FIXC1*inverseClock
Clock [MHz] 1.E-06*(FIXC1/FIXC2)/inverseClock
CPI FIXC1/FIXC0
Local LLC hit with false sharing [MByte] 1.E-06*PMC0*64
Local LLC hit with false sharing rate PMC0/PMC2
Remote LLC false sharing [MByte] 1.E-06*PMC1*64
Remote LLC false sharing rate PMC1/PMC2
LONG
Formulas:
Local LLC false sharing [MByte] = 1.E-06*MEM_LOAD_UOPS_L3_HIT_RETIRED_XSNP_HITM*64
Local LLC false sharing rate = MEM_LOAD_UOPS_L3_HIT_RETIRED_XSNP_HITM/MEM_LOAD_UOPS_RETIRED_ALL
Remote LLC false sharing [MByte] = 1.E-06*MEM_LOAD_UOPS_L3_MISS_RETIRED_REMOTE_HITM*64
Remote LLC false sharing rate = MEM_LOAD_UOPS_L3_MISS_RETIRED_REMOTE_HITM/MEM_LOAD_UOPS_RETIRED_ALL
-
False-sharing of cache lines can dramatically reduce the performance of an
application. This performance group measures the L3 traffic induced by false-sharing.
The false-sharing rate uses all memory loads as reference.
For systems with multiple CPU sockets, this performance group also measures the
false-sharing of cache lines over socket boundaries.
Please keep in mind that the MEM_LOAD_UOPS_L3_HIT_RETIRED_XSNP_HITM event may
undercount by as much as 40% (Errata HSW150).