1 Started Nov 1999 by Kanoj Sarcar <kanoj@sgi.com>
5 This question can be answered from a couple of perspectives: the
6 hardware view and the Linux software view.
8 From the hardware perspective, a NUMA system is a computer platform that
9 comprises multiple components or assemblies each of which may contain 0
10 or more CPUs, local memory, and/or IO buses. For brevity and to
11 disambiguate the hardware view of these physical components/assemblies
12 from the software abstraction thereof, we'll call the components/assemblies
13 'cells' in this document.
15 Each of the 'cells' may be viewed as an SMP [symmetric multi-processor] subset
16 of the system--although some components necessary for a stand-alone SMP system
17 may not be populated on any given cell. The cells of the NUMA system are
18 connected together with some sort of system interconnect--e.g., a crossbar or
19 point-to-point link are common types of NUMA system interconnects. Both of
20 these types of interconnects can be aggregated to create NUMA platforms with
21 cells at multiple distances from other cells.
23 For Linux, the NUMA platforms of interest are primarily what is known as Cache
24 Coherent NUMA or ccNUMA systems. With ccNUMA systems, all memory is visible
25 to and accessible from any CPU attached to any cell and cache coherency
26 is handled in hardware by the processor caches and/or the system interconnect.
28 Memory access time and effective memory bandwidth varies depending on how far
29 away the cell containing the CPU or IO bus making the memory access is from the
30 cell containing the target memory. For example, access to memory by CPUs
31 attached to the same cell will experience faster access times and higher
32 bandwidths than accesses to memory on other, remote cells. NUMA platforms
33 can have cells at multiple remote distances from any given cell.
35 Platform vendors don't build NUMA systems just to make software developers'
36 lives interesting. Rather, this architecture is a means to provide scalable
37 memory bandwidth. However, to achieve scalable memory bandwidth, system and
38 application software must arrange for a large majority of the memory references
39 [cache misses] to be to "local" memory--memory on the same cell, if any--or
40 to the closest cell with memory.
42 This leads to the Linux software view of a NUMA system:
44 Linux divides the system's hardware resources into multiple software
45 abstractions called "nodes". Linux maps the nodes onto the physical cells
46 of the hardware platform, abstracting away some of the details for some
47 architectures. As with physical cells, software nodes may contain 0 or more
48 CPUs, memory and/or IO buses. And, again, memory accesses to memory on
49 "closer" nodes--nodes that map to closer cells--will generally experience
50 faster access times and higher effective bandwidth than accesses to more
53 For some architectures, such as x86, Linux will "hide" any node representing a
54 physical cell that has no memory attached, and reassign any CPUs attached to
55 that cell to a node representing a cell that does have memory. Thus, on
56 these architectures, one cannot assume that all CPUs that Linux associates with
57 a given node will see the same local memory access times and bandwidth.
59 In addition, for some architectures, again x86 is an example, Linux supports
60 the emulation of additional nodes. For NUMA emulation, linux will carve up
61 the existing nodes--or the system memory for non-NUMA platforms--into multiple
62 nodes. Each emulated node will manage a fraction of the underlying cells'
63 physical memory. NUMA emluation is useful for testing NUMA kernel and
64 application features on non-NUMA platforms, and as a sort of memory resource
65 management mechanism when used together with cpusets.
66 [see Documentation/cgroup-v1/cpusets.txt]
68 For each node with memory, Linux constructs an independent memory management
69 subsystem, complete with its own free page lists, in-use page lists, usage
70 statistics and locks to mediate access. In addition, Linux constructs for
71 each memory zone [one or more of DMA, DMA32, NORMAL, HIGH_MEMORY, MOVABLE],
72 an ordered "zonelist". A zonelist specifies the zones/nodes to visit when a
73 selected zone/node cannot satisfy the allocation request. This situation,
74 when a zone has no available memory to satisfy a request, is called
75 "overflow" or "fallback".
77 Because some nodes contain multiple zones containing different types of
78 memory, Linux must decide whether to order the zonelists such that allocations
79 fall back to the same zone type on a different node, or to a different zone
80 type on the same node. This is an important consideration because some zones,
81 such as DMA or DMA32, represent relatively scarce resources. Linux chooses
82 a default zonelist order based on the sizes of the various zone types relative
83 to the total memory of the node and the total memory of the system. The
84 default zonelist order may be overridden using the numa_zonelist_order kernel
85 boot parameter or sysctl. [see Documentation/kernel-parameters.txt and
86 Documentation/sysctl/vm.txt]
88 By default, Linux will attempt to satisfy memory allocation requests from the
89 node to which the CPU that executes the request is assigned. Specifically,
90 Linux will attempt to allocate from the first node in the appropriate zonelist
91 for the node where the request originates. This is called "local allocation."
92 If the "local" node cannot satisfy the request, the kernel will examine other
93 nodes' zones in the selected zonelist looking for the first zone in the list
94 that can satisfy the request.
96 Local allocation will tend to keep subsequent access to the allocated memory
97 "local" to the underlying physical resources and off the system interconnect--
98 as long as the task on whose behalf the kernel allocated some memory does not
99 later migrate away from that memory. The Linux scheduler is aware of the
100 NUMA topology of the platform--embodied in the "scheduling domains" data
101 structures [see Documentation/scheduler/sched-domains.txt]--and the scheduler
102 attempts to minimize task migration to distant scheduling domains. However,
103 the scheduler does not take a task's NUMA footprint into account directly.
104 Thus, under sufficient imbalance, tasks can migrate between nodes, remote
105 from their initial node and kernel data structures.
107 System administrators and application designers can restrict a task's migration
108 to improve NUMA locality using various CPU affinity command line interfaces,
109 such as taskset(1) and numactl(1), and program interfaces such as
110 sched_setaffinity(2). Further, one can modify the kernel's default local
111 allocation behavior using Linux NUMA memory policy.
112 [see Documentation/vm/numa_memory_policy.txt.]
114 System administrators can restrict the CPUs and nodes' memories that a non-
115 privileged user can specify in the scheduling or NUMA commands and functions
116 using control groups and CPUsets. [see Documentation/cgroup-v1/cpusets.txt]
118 On architectures that do not hide memoryless nodes, Linux will include only
119 zones [nodes] with memory in the zonelists. This means that for a memoryless
120 node the "local memory node"--the node of the first zone in CPU's node's
121 zonelist--will not be the node itself. Rather, it will be the node that the
122 kernel selected as the nearest node with memory when it built the zonelists.
123 So, default, local allocations will succeed with the kernel supplying the
124 closest available memory. This is a consequence of the same mechanism that
125 allows such allocations to fallback to other nearby nodes when a node that
126 does contain memory overflows.
128 Some kernel allocations do not want or cannot tolerate this allocation fallback
129 behavior. Rather they want to be sure they get memory from the specified node
130 or get notified that the node has no free memory. This is usually the case when
131 a subsystem allocates per CPU memory resources, for example.
133 A typical model for making such an allocation is to obtain the node id of the
134 node to which the "current CPU" is attached using one of the kernel's
135 numa_node_id() or CPU_to_node() functions and then request memory from only
136 the node id returned. When such an allocation fails, the requesting subsystem
137 may revert to its own fallback path. The slab kernel memory allocator is an
138 example of this. Or, the subsystem may choose to disable or not to enable
139 itself on allocation failure. The kernel profiling subsystem is an example of
142 If the architecture supports--does not hide--memoryless nodes, then CPUs
143 attached to memoryless nodes would always incur the fallback path overhead
144 or some subsystems would fail to initialize if they attempted to allocated
145 memory exclusively from a node without memory. To support such
146 architectures transparently, kernel subsystems can use the numa_mem_id()
147 or cpu_to_mem() function to locate the "local memory node" for the calling or
148 specified CPU. Again, this is the same node from which default, local page
149 allocations will be attempted.