1 perf-amd-ibs(1)
2 ===============
3
4 NAME
5 ----
6 perf-amd-ibs - Support for AMD Instruction-Bas
7
8 SYNOPSIS
9 --------
10 [verse]
11 'perf record' -e ibs_op//
12 'perf record' -e ibs_fetch//
13
14 DESCRIPTION
15 -----------
16
17 Instruction-Based Sampling (IBS) provides prec
18 profiling support on AMD platforms. IBS has tw
19 Op and IBS Fetch. IBS Op sampling provides inf
20 execution (micro-op execution to be precise) w
21 hit/miss, d-TLB hit/miss, cache miss latency,
22 behavior etc. IBS Fetch sampling provides info
23 with details like i-cache hit/miss, i-TLB hit/
24 per-smt-thread i.e. each SMT hardware thread c
25
26 Both, IBS Op and IBS Fetch, are exposed as PMU
27 using the Linux perf utility. The following fi
28 if IBS is supported by the hardware and kernel
29
30 /sys/bus/event_source/devices/ibs_op/
31 /sys/bus/event_source/devices/ibs_fetch/
32
33 IBS Op PMU supports two events: cycles and mic
34 one event: fetch ops.
35
36 IBS PMUs do not have user/kernel filtering cap
37 CAP_SYS_ADMIN or CAP_PERFMON privilege.
38
39 IBS VS. REGULAR CORE PMU
40 ------------------------
41
42 IBS gives samples with precise IP, i.e. the IP
43 no skid. Whereas the IP recorded by regular co
44 (sample was generated at IP X but perf would r
45 regular core PMU might not help for profiling
46 precision. Further, IBS provides additional in
47 question. On the other hand, regular core PMU
48 plethora of events, counting mode (less interf
49 counters, event grouping support, filtering ca
50
51 Three regular core PMU events are internally f
52 precise_ip attribute is set:
53
54 -e cpu-cycles:p becomes -e ibs_op//
55 -e r076:p becomes -e ibs_op//
56 -e r0C1:p becomes -e ibs_op/cnt_ctl=1/
57
58 EXAMPLES
59 --------
60
61 IBS Op PMU
62 ~~~~~~~~~~
63
64 System-wide profile, cycles event, sampling pe
65
66 # perf record -e ibs_op// -c 100000 -a
67
68 Per-cpu profile (cpu10), cycles event, samplin
69
70 # perf record -e ibs_op// -c 100000 -C
71
72 Per-cpu profile (cpu10), cycles event, samplin
73
74 # perf record -e ibs_op// -F 1000 -C 1
75
76 System-wide profile, uOps event, sampling peri
77
78 # perf record -e ibs_op/cnt_ctl=1/ -c
79
80 Same command, but also capture IBS register ra
81
82 # perf record -e ibs_op/cnt_ctl=1/ -c
83
84 System-wide profile, uOps event, sampling peri
85
86 # perf record -e ibs_op/cnt_ctl=1,l3mi
87
88 Per process(upstream v6.2 onward), uOps event,
89
90 # perf record -e ibs_op/cnt_ctl=1/ -c
91
92 Per process(upstream v6.2 onward), uOps event,
93
94 # perf record -e ibs_op/cnt_ctl=1/ -c
95
96 To analyse recorded profile in aggregate mode
97
98 # perf report
99 /* Select a line and press 'a' to dril
100
101 To go over each sample
102
103 # perf script
104
105 Raw dump of IBS registers when profiled with -
106
107 # perf report -D
108 /* Look for PERF_RECORD_SAMPLE */
109
110 Example register raw dump:
111
112 ibs_op_ctl: 000002c30006186a MaxCn
113 Val 1 CntCtl 0=cycles CurCnt
114 IbsOpRip: ffffffff8204aea7
115 ibs_op_data: 0000010002550001 CompT
116 BrnRet 0 RipInvalid 0 BrnFuse
117 ibs_op_data2: 0000000000000013 RmtNo
118 ibs_op_data3: 0000000031960092 LdOp
119 DcL2TlbMiss 0 DcL1TlbHit2M 1 D
120 DcMiss 1 DcMisAcc 0 DcWcMemAcc
121 DcMissNoMabAlloc 0 DcLinAddrVa
122 DcL2TlbHit1G 0 L2Miss 1 SwPf 0
123 OpDcMissOpenMemReqs 12 DcMissL
124 IbsDCLinAd: ff110008a5398920
125 IbsDCPhysAd: 00000008a5398920
126
127 IBS applied in a real world usecase
128
129 ~90% regression was observed in tbench
130 which was counter intuitive. IBS profi
131 using perf helped in identifying exact
132
133 https://lore.kernel.org/r/202209210636
134
135 IBS Fetch PMU
136 ~~~~~~~~~~~~~
137
138 Similar commands can be used with Fetch PMU as
139
140 System-wide profile, fetch ops event, sampling
141
142 # perf record -e ibs_fetch// -c 100000
143
144 System-wide profile, fetch ops event, sampling
145
146 # perf record -e ibs_fetch/rand_en=1/
147
148 Random enable adds small degree of var
149 helps in cases like long running loops
150 instruction over and over because of f
151
152 etc.
153
154 PERF MEM AND PERF C2C
155 ---------------------
156
157 perf mem is a memory access profiler tool and
158 cacheline analyser tool. Both of them internal
159 Below is a simple example of the perf mem tool
160
161 # perf mem record -c 100000 -- make
162 # perf mem report
163
164 A normal perf mem report output will provide d
165 However, it can also be aggregated based on ou
166
167 # perf mem report -F mem,sample,snoop
168 Samples: 3M of event 'ibs_op//', Event
169 Memory access
170 N/A
171 L1 hit
172 L2 hit
173 L3 hit
174 L3 hit
175 RAM hit
176 Remote node, same socket RAM hit
177 Remote core, same node Any cache hit
178 Remote core, same node Any cache hit
179 Remote node, same socket Any cache hit
180 Remote node, same socket Any cache hit
181 Uncached hit
182
183 Please refer to their man page for more detail
184
185 SEE ALSO
186 --------
187
188 linkperf:perf-record[1], linkperf:perf-script[
189 linkperf:perf-mem[1], linkperf:perf-c2c[1]
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.