Loading...
Note: File does not exist in v3.1.
1Intel hybrid support
2--------------------
3Support for Intel hybrid events within perf tools.
4
5For some Intel platforms, such as AlderLake, which is hybrid platform and
6it consists of atom cpu and core cpu. Each cpu has dedicated event list.
7Part of events are available on core cpu, part of events are available
8on atom cpu and even part of events are available on both.
9
10Kernel exports two new cpu pmus via sysfs:
11/sys/devices/cpu_core
12/sys/devices/cpu_atom
13
14The 'cpus' files are created under the directories. For example,
15
16cat /sys/devices/cpu_core/cpus
170-15
18
19cat /sys/devices/cpu_atom/cpus
2016-23
21
22It indicates cpu0-cpu15 are core cpus and cpu16-cpu23 are atom cpus.
23
24Quickstart
25
26List hybrid event
27-----------------
28
29As before, use perf-list to list the symbolic event.
30
31perf list
32
33inst_retired.any
34 [Fixed Counter: Counts the number of instructions retired. Unit: cpu_atom]
35inst_retired.any
36 [Number of instructions retired. Fixed Counter - architectural event. Unit: cpu_core]
37
38The 'Unit: xxx' is added to brief description to indicate which pmu
39the event is belong to. Same event name but with different pmu can
40be supported.
41
42Enable hybrid event with a specific pmu
43---------------------------------------
44
45To enable a core only event or atom only event, following syntax is supported:
46
47 cpu_core/<event name>/
48or
49 cpu_atom/<event name>/
50
51For example, count the 'cycles' event on core cpus.
52
53 perf stat -e cpu_core/cycles/
54
55Create two events for one hardware event automatically
56------------------------------------------------------
57
58When creating one event and the event is available on both atom and core,
59two events are created automatically. One is for atom, the other is for
60core. Most of hardware events and cache events are available on both
61cpu_core and cpu_atom.
62
63For hardware events, they have pre-defined configs (e.g. 0 for cycles).
64But on hybrid platform, kernel needs to know where the event comes from
65(from atom or from core). The original perf event type PERF_TYPE_HARDWARE
66can't carry pmu information. So now this type is extended to be PMU aware
67type. The PMU type ID is stored at attr.config[63:32].
68
69PMU type ID is retrieved from sysfs.
70/sys/devices/cpu_atom/type
71/sys/devices/cpu_core/type
72
73The new attr.config layout for PERF_TYPE_HARDWARE:
74
75PERF_TYPE_HARDWARE: 0xEEEEEEEE000000AA
76 AA: hardware event ID
77 EEEEEEEE: PMU type ID
78
79Cache event is similar. The type PERF_TYPE_HW_CACHE is extended to be
80PMU aware type. The PMU type ID is stored at attr.config[63:32].
81
82The new attr.config layout for PERF_TYPE_HW_CACHE:
83
84PERF_TYPE_HW_CACHE: 0xEEEEEEEE00DDCCBB
85 BB: hardware cache ID
86 CC: hardware cache op ID
87 DD: hardware cache op result ID
88 EEEEEEEE: PMU type ID
89
90When enabling a hardware event without specified pmu, such as,
91perf stat -e cycles -a (use system-wide in this example), two events
92are created automatically.
93
94 ------------------------------------------------------------
95 perf_event_attr:
96 size 120
97 config 0x400000000
98 sample_type IDENTIFIER
99 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
100 disabled 1
101 inherit 1
102 exclude_guest 1
103 ------------------------------------------------------------
104
105and
106
107 ------------------------------------------------------------
108 perf_event_attr:
109 size 120
110 config 0x800000000
111 sample_type IDENTIFIER
112 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
113 disabled 1
114 inherit 1
115 exclude_guest 1
116 ------------------------------------------------------------
117
118type 0 is PERF_TYPE_HARDWARE.
1190x4 in 0x400000000 indicates it's cpu_core pmu.
1200x8 in 0x800000000 indicates it's cpu_atom pmu (atom pmu type id is random).
121
122The kernel creates 'cycles' (0x400000000) on cpu0-cpu15 (core cpus),
123and create 'cycles' (0x800000000) on cpu16-cpu23 (atom cpus).
124
125For perf-stat result, it displays two events:
126
127 Performance counter stats for 'system wide':
128
129 6,744,979 cpu_core/cycles/
130 1,965,552 cpu_atom/cycles/
131
132The first 'cycles' is core event, the second 'cycles' is atom event.
133
134Thread mode example:
135--------------------
136
137perf-stat reports the scaled counts for hybrid event and with a percentage
138displayed. The percentage is the event's running time/enabling time.
139
140One example, 'triad_loop' runs on cpu16 (atom core), while we can see the
141scaled value for core cycles is 160,444,092 and the percentage is 0.47%.
142
143perf stat -e cycles -- taskset -c 16 ./triad_loop
144
145As previous, two events are created.
146
147------------------------------------------------------------
148perf_event_attr:
149 size 120
150 config 0x400000000
151 sample_type IDENTIFIER
152 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
153 disabled 1
154 inherit 1
155 enable_on_exec 1
156 exclude_guest 1
157------------------------------------------------------------
158
159and
160
161------------------------------------------------------------
162perf_event_attr:
163 size 120
164 config 0x800000000
165 sample_type IDENTIFIER
166 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
167 disabled 1
168 inherit 1
169 enable_on_exec 1
170 exclude_guest 1
171------------------------------------------------------------
172
173 Performance counter stats for 'taskset -c 16 ./triad_loop':
174
175 233,066,666 cpu_core/cycles/ (0.43%)
176 604,097,080 cpu_atom/cycles/ (99.57%)
177
178perf-record:
179------------
180
181If there is no '-e' specified in perf record, on hybrid platform,
182it creates two default 'cycles' and adds them to event list. One
183is for core, the other is for atom.
184
185perf-stat:
186----------
187
188If there is no '-e' specified in perf stat, on hybrid platform,
189besides of software events, following events are created and
190added to event list in order.
191
192cpu_core/cycles/,
193cpu_atom/cycles/,
194cpu_core/instructions/,
195cpu_atom/instructions/,
196cpu_core/branches/,
197cpu_atom/branches/,
198cpu_core/branch-misses/,
199cpu_atom/branch-misses/
200
201Of course, both perf-stat and perf-record support to enable
202hybrid event with a specific pmu.
203
204e.g.
205perf stat -e cpu_core/cycles/
206perf stat -e cpu_atom/cycles/
207perf stat -e cpu_core/r1a/
208perf stat -e cpu_atom/L1-icache-loads/
209perf stat -e cpu_core/cycles/,cpu_atom/instructions/
210perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}'
211
212But '{cpu_core/cycles/,cpu_atom/instructions/}' will return
213warning and disable grouping, because the pmus in group are
214not matched (cpu_core vs. cpu_atom).