1 ======================= !! 1 .. _numaperf:
2 NUMA Memory Performance <<
3 ======================= <<
4 2
>> 3 =============
5 NUMA Locality 4 NUMA Locality
6 ============= 5 =============
7 6
8 Some platforms may have multiple types of memo 7 Some platforms may have multiple types of memory attached to a compute
9 node. These disparate memory ranges may share 8 node. These disparate memory ranges may share some characteristics, such
10 as CPU cache coherence, but may have different 9 as CPU cache coherence, but may have different performance. For example,
11 different media types and buses affect bandwid 10 different media types and buses affect bandwidth and latency.
12 11
13 A system supports such heterogeneous memory by 12 A system supports such heterogeneous memory by grouping each memory type
14 under different domains, or "nodes", based on 13 under different domains, or "nodes", based on locality and performance
15 characteristics. Some memory may share the sa 14 characteristics. Some memory may share the same node as a CPU, and others
16 are provided as memory only nodes. While memor 15 are provided as memory only nodes. While memory only nodes do not provide
17 CPUs, they may still be local to one or more c 16 CPUs, they may still be local to one or more compute nodes relative to
18 other nodes. The following diagram shows one s 17 other nodes. The following diagram shows one such example of two compute
19 nodes with local memory and a memory only node 18 nodes with local memory and a memory only node for each of compute node::
20 19
21 +------------------+ +------------------+ 20 +------------------+ +------------------+
22 | Compute Node 0 +-----+ Compute Node 1 | 21 | Compute Node 0 +-----+ Compute Node 1 |
23 | Local Node0 Mem | | Local Node1 Mem | 22 | Local Node0 Mem | | Local Node1 Mem |
24 +--------+---------+ +--------+---------+ 23 +--------+---------+ +--------+---------+
25 | | 24 | |
26 +--------+---------+ +--------+---------+ 25 +--------+---------+ +--------+---------+
27 | Slower Node2 Mem | | Slower Node3 Mem | 26 | Slower Node2 Mem | | Slower Node3 Mem |
28 +------------------+ +--------+---------+ 27 +------------------+ +--------+---------+
29 28
30 A "memory initiator" is a node containing one 29 A "memory initiator" is a node containing one or more devices such as
31 CPUs or separate memory I/O devices that can i 30 CPUs or separate memory I/O devices that can initiate memory requests.
32 A "memory target" is a node containing one or 31 A "memory target" is a node containing one or more physical address
33 ranges accessible from one or more memory init 32 ranges accessible from one or more memory initiators.
34 33
35 When multiple memory initiators exist, they ma 34 When multiple memory initiators exist, they may not all have the same
36 performance when accessing a given memory targ 35 performance when accessing a given memory target. Each initiator-target
37 pair may be organized into different ranked ac 36 pair may be organized into different ranked access classes to represent
38 this relationship. The highest performing init 37 this relationship. The highest performing initiator to a given target
39 is considered to be one of that target's local 38 is considered to be one of that target's local initiators, and given
40 the highest access class, 0. Any given target 39 the highest access class, 0. Any given target may have one or more
41 local initiators, and any given initiator may 40 local initiators, and any given initiator may have multiple local
42 memory targets. 41 memory targets.
43 42
44 To aid applications matching memory targets wi 43 To aid applications matching memory targets with their initiators, the
45 kernel provides symlinks to each other. The fo 44 kernel provides symlinks to each other. The following example lists the
46 relationship for the access class "0" memory i 45 relationship for the access class "0" memory initiators and targets::
47 46
48 # symlinks -v /sys/devices/system/node 47 # symlinks -v /sys/devices/system/node/nodeX/access0/targets/
49 relative: /sys/devices/system/node/nod 48 relative: /sys/devices/system/node/nodeX/access0/targets/nodeY -> ../../nodeY
50 49
51 # symlinks -v /sys/devices/system/node 50 # symlinks -v /sys/devices/system/node/nodeY/access0/initiators/
52 relative: /sys/devices/system/node/nod 51 relative: /sys/devices/system/node/nodeY/access0/initiators/nodeX -> ../../nodeX
53 52
54 A memory initiator may have multiple memory ta 53 A memory initiator may have multiple memory targets in the same access
55 class. The target memory's initiators in a giv 54 class. The target memory's initiators in a given class indicate the
56 nodes' access characteristics share the same p 55 nodes' access characteristics share the same performance relative to other
57 linked initiator nodes. Each target within an 56 linked initiator nodes. Each target within an initiator's access class,
58 though, do not necessarily perform the same as 57 though, do not necessarily perform the same as each other.
59 58
60 The access class "1" is used to allow differen !! 59 ================
61 that are CPUs and hence suitable for generic t <<
62 IO initiators such as GPUs and NICs. Unlike a <<
63 nodes containing CPUs are considered. <<
64 <<
65 NUMA Performance 60 NUMA Performance
66 ================ 61 ================
67 62
68 Applications may wish to consider which node t 63 Applications may wish to consider which node they want their memory to
69 be allocated from based on the node's performa 64 be allocated from based on the node's performance characteristics. If
70 the system provides these attributes, the kern 65 the system provides these attributes, the kernel exports them under the
71 node sysfs hierarchy by appending the attribut 66 node sysfs hierarchy by appending the attributes directory under the
72 memory node's access class 0 initiators as fol 67 memory node's access class 0 initiators as follows::
73 68
74 /sys/devices/system/node/nodeY/access0 69 /sys/devices/system/node/nodeY/access0/initiators/
75 70
76 These attributes apply only when accessed from 71 These attributes apply only when accessed from nodes that have the
77 are linked under the this access's initiators. !! 72 are linked under the this access's inititiators.
78 73
79 The performance characteristics the kernel pro 74 The performance characteristics the kernel provides for the local initiators
80 are exported are as follows:: 75 are exported are as follows::
81 76
82 # tree -P "read*|write*" /sys/devices/ 77 # tree -P "read*|write*" /sys/devices/system/node/nodeY/access0/initiators/
83 /sys/devices/system/node/nodeY/access0 78 /sys/devices/system/node/nodeY/access0/initiators/
84 |-- read_bandwidth 79 |-- read_bandwidth
85 |-- read_latency 80 |-- read_latency
86 |-- write_bandwidth 81 |-- write_bandwidth
87 `-- write_latency 82 `-- write_latency
88 83
89 The bandwidth attributes are provided in MiB/s 84 The bandwidth attributes are provided in MiB/second.
90 85
91 The latency attributes are provided in nanosec 86 The latency attributes are provided in nanoseconds.
92 87
93 The values reported here correspond to the rat 88 The values reported here correspond to the rated latency and bandwidth
94 for the platform. 89 for the platform.
95 90
96 Access class 1 takes the same form but only in !! 91 ==========
97 memory activity. <<
98 <<
99 NUMA Cache 92 NUMA Cache
100 ========== 93 ==========
101 94
102 System memory may be constructed in a hierarch 95 System memory may be constructed in a hierarchy of elements with various
103 performance characteristics in order to provid 96 performance characteristics in order to provide large address space of
104 slower performing memory cached by a smaller h 97 slower performing memory cached by a smaller higher performing memory. The
105 system physical addresses memory initiators a 98 system physical addresses memory initiators are aware of are provided
106 by the last memory level in the hierarchy. The 99 by the last memory level in the hierarchy. The system meanwhile uses
107 higher performing memory to transparently cach 100 higher performing memory to transparently cache access to progressively
108 slower levels. 101 slower levels.
109 102
110 The term "far memory" is used to denote the la 103 The term "far memory" is used to denote the last level memory in the
111 hierarchy. Each increasing cache level provide 104 hierarchy. Each increasing cache level provides higher performing
112 initiator access, and the term "near memory" r 105 initiator access, and the term "near memory" represents the fastest
113 cache provided by the system. 106 cache provided by the system.
114 107
115 This numbering is different than CPU caches wh 108 This numbering is different than CPU caches where the cache level (ex:
116 L1, L2, L3) uses the CPU-side view where each 109 L1, L2, L3) uses the CPU-side view where each increased level is lower
117 performing. In contrast, the memory cache leve 110 performing. In contrast, the memory cache level is centric to the last
118 level memory, so the higher numbered cache lev 111 level memory, so the higher numbered cache level corresponds to memory
119 nearer to the CPU, and further from far memory 112 nearer to the CPU, and further from far memory.
120 113
121 The memory-side caches are not directly addres 114 The memory-side caches are not directly addressable by software. When
122 software accesses a system address, the system 115 software accesses a system address, the system will return it from the
123 near memory cache if it is present. If it is n 116 near memory cache if it is present. If it is not present, the system
124 accesses the next level of memory until there 117 accesses the next level of memory until there is either a hit in that
125 cache level, or it reaches far memory. 118 cache level, or it reaches far memory.
126 119
127 An application does not need to know about cac 120 An application does not need to know about caching attributes in order
128 to use the system. Software may optionally que 121 to use the system. Software may optionally query the memory cache
129 attributes in order to maximize the performanc 122 attributes in order to maximize the performance out of such a setup.
130 If the system provides a way for the kernel to 123 If the system provides a way for the kernel to discover this information,
131 for example with ACPI HMAT (Heterogeneous Memo 124 for example with ACPI HMAT (Heterogeneous Memory Attribute Table),
132 the kernel will append these attributes to the 125 the kernel will append these attributes to the NUMA node memory target.
133 126
134 When the kernel first registers a memory cache 127 When the kernel first registers a memory cache with a node, the kernel
135 will create the following directory:: 128 will create the following directory::
136 129
137 /sys/devices/system/node/nodeX/memory_ 130 /sys/devices/system/node/nodeX/memory_side_cache/
138 131
139 If that directory is not present, the system e 132 If that directory is not present, the system either does not provide
140 a memory-side cache, or that information is no 133 a memory-side cache, or that information is not accessible to the kernel.
141 134
142 The attributes for each level of cache is prov 135 The attributes for each level of cache is provided under its cache
143 level index:: 136 level index::
144 137
145 /sys/devices/system/node/nodeX/memory_ 138 /sys/devices/system/node/nodeX/memory_side_cache/indexA/
146 /sys/devices/system/node/nodeX/memory_ 139 /sys/devices/system/node/nodeX/memory_side_cache/indexB/
147 /sys/devices/system/node/nodeX/memory_ 140 /sys/devices/system/node/nodeX/memory_side_cache/indexC/
148 141
149 Each cache level's directory provides its attr 142 Each cache level's directory provides its attributes. For example, the
150 following shows a single cache level and the a 143 following shows a single cache level and the attributes available for
151 software to query:: 144 software to query::
152 145
153 # tree /sys/devices/system/node/node0/ !! 146 # tree sys/devices/system/node/node0/memory_side_cache/
154 /sys/devices/system/node/node0/memory_ 147 /sys/devices/system/node/node0/memory_side_cache/
155 |-- index1 148 |-- index1
156 | |-- indexing 149 | |-- indexing
157 | |-- line_size 150 | |-- line_size
158 | |-- size 151 | |-- size
159 | `-- write_policy 152 | `-- write_policy
160 153
161 The "indexing" will be 0 if it is a direct-map 154 The "indexing" will be 0 if it is a direct-mapped cache, and non-zero
162 for any other indexed based, multi-way associa 155 for any other indexed based, multi-way associativity.
163 156
164 The "line_size" is the number of bytes accesse 157 The "line_size" is the number of bytes accessed from the next cache
165 level on a miss. 158 level on a miss.
166 159
167 The "size" is the number of bytes provided by 160 The "size" is the number of bytes provided by this cache level.
168 161
169 The "write_policy" will be 0 for write-back, a 162 The "write_policy" will be 0 for write-back, and non-zero for
170 write-through caching. 163 write-through caching.
171 164
>> 165 ========
172 See Also 166 See Also
173 ======== 167 ========
174 168
175 [1] https://www.uefi.org/sites/default/files/r 169 [1] https://www.uefi.org/sites/default/files/resources/ACPI_6_2.pdf
176 - Section 5.2.27 170 - Section 5.2.27
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.