1LIBPFM(3) Linux Programmer's Manual LIBPFM(3)
2
6 libpfm_intel_knl - support for Intel Kinghts Landing core PMU
7
9 #include <perfmon/pfmlib.h>
10 PMU name: knl
12 PMU desc: Intel Kinghts Landing
13
16 The library supports the Intel Kinghts Landing core PMU. It should be
17 noted that this PMU model only covers each core's PMU and not the
18 socket level PMU.
19 On Knights Landing, the number of generic counters is 4. There is 4-way
21 HyperThreading support. The pfm_get_pmu_info() function returns the
22 maximum number of generic counters in num_cntrs.
23
26 The following modifiers are supported on Intel Kinghts Landing proces‐
27 sors:
28 u Measure at user level which includes privilege levels 1, 2, 3.
30 This corresponds to PFM_PLM3. This is a boolean modifier.
31 k Measure at kernel level which includes privilege level 0. This
33 corresponds to PFM_PLM0. This is a boolean modifier.
34 i Invert the meaning of the event. The counter will now count
36 cycles in which the event is not occurring. This is a boolean
37 modifier
38 e Enable edge detection, i.e., count only when there is a state
40 transition from no occurrence of the event to at least one
41 occurrence. This modifier must be combined with a counter mask
42 modifier (m) with a value greater or equal to one. This is a
43 boolean modifier.
44 c Set the counter mask value. The mask acts as a threshold. The
46 counter will count the number of cycles in which the number of
47 occurrences of the event is greater or equal to the threshold.
48 This is an integer modifier with values in the range [0:255].
49 t Measure on any of the 4 hyper-threads at the same time assuming
51 hyper-threading is enabled. This is a boolean modifier. This
52 modifier is only available on fixed counters (unhalted_refer‐
53 ence_cycles, instructions_retired, unhalted_core_cycles).
54 Depending on the underlying kernel interface, the event may be
55 programmed on a fixed counter or a generic counter, except for
56 unhalted_reference_cycles, in which case, this modifier may be
57 ignored or rejected.
58
61 Intel Knights Landing provides two offcore_response events. They are
62 called OFFCORE_RESPONSE_0 and OFFCORE_RESPONSE_1.
63 Those events need special treatment in the performance monitoring in‐
65 frastructure because each event uses an extra register to store some
66 settings. Thus, in case multiple offcore_response events are monitored
67 simultaneously, the kernel needs to manage the sharing of that extra
68 register.
69 The offcore_response events are exposed as normal events by the
71 library. The extra settings are exposed as regular umasks. The library
72 takes care of encoding the events according to the underlying kernel
73 interface.
74 On Intel Knights Landing, the umasks are divided into 4 categories:
76 request, supplier and snoop and average latency. Offcore_response event
77 has two modes of operations: normal and average latency. In the first
78 mode, the two offcore_respnse events operate independently of each
79 other. The user must provide at least one umask for each of the first 3
80 categories: request, supplier, snoop. In the second mode, the two off‐
81 core_response events are combined to compute an average latency per
82 request type.
83 For the normal mode, there is a special supplier (response) umask
85 called ANY_RESPONSE. When this umask is used then it overrides any sup‐
86 plier and snoop umasks. In other words, users can specify either
87 ANY_RESPONSE OR any combinations of supplier + snoops. In case no sup‐
88 plier or snoop is specified, the library defaults to using
89 ANY_RESPONSE.
90 For instance, the following are valid event selections:
92 OFFCORE_RESPONSE_0:DMND_DATA_RD:ANY_RESPONSE
94 OFFCORE_RESPONSE_0:ANY_REQUEST
96 OFFCORE_RESPONSE_0:ANY_RFO:DDR_NEAR
98 But the following is illegal:
101 OFFCORE_RESPONSE_0:ANY_RFO:DDR_NEAR:ANY_RESPONSE
104 In average latency mode, OFFCORE_RESPONSE_0 must be programmed to
106 select the request types of interest, for instance, DMND_DATA_RD, and
107 the OUTSTANDING umask must be set and no others. the library will
108 enforce that restriction as soon as the OUTSTANDING umask is used. Then
109 OFFCORE_RESPONSE_1 must be set with the same request types and the
110 ANY_RESPONSE umask. It should be noted that the library encodes events
111 independently of each other and therefore cannot verify that the
112 requests are matching between the two events. Example of average
113 latency settings:
114 OFFCORE_RESPONSE_0:DMND_DATA_RD:OUTSTANDING+OFF‐
116 CORE_RESPONSE_1:DMND_DATA_RD:ANY_RESPONSE
117 OFFCORE_RESPONSE_0:ANY_REQUEST:OUTSTANDING+OFF‐
119 CORE_RESPONSE_1:ANY_REQUEST:ANY_RESPONSE
120 The average latency for the request(s) is obtained by dividing the
122 counts of OFFCORE_RESPONSE_0 by the count of OFFCORE_RESPONSE_1. The
123 ratio is expressed in core cycles.
124
127 Stephane Eranian <eranian@gmail.com>
128 July, 2016 LIBPFM(3)