1 Started Nov 1999 by Kanoj Sarcar <kanoj@sgi.com> 2 3 ============= 4 What is NUMA? 5 ============= 6 7 This question can be answered from a couple of perspectives: the 8 hardware view and the Linux software view. 9 10 From the hardware perspective, a NUMA system is a computer platform that 11 comprises multiple components or assemblies each of which may contain 0 12 or more CPUs, local memory, and/or IO buses. For brevity and to 13 disambiguate the hardware view of these physical components/assemblies 14 from the software abstraction thereof, we'll call the components/assemblies 15 'cells' in this document. 16 17 Each of the 'cells' may be viewed as an SMP [symmetric multi-processor] subset 18 of the system--although some components necessary for a stand-alone SMP system 19 may not be populated on any given cell. The cells of the NUMA system are 20 connected together with some sort of system interconnect--e.g., a crossbar or 21 point-to-point link are common types of NUMA system interconnects. Both of 22 these types of interconnects can be aggregated to create NUMA platforms with 23 cells at multiple distances from other cells. 24 25 For Linux, the NUMA platforms of interest are primarily what is known as Cache 26 Coherent NUMA or ccNUMA systems. With ccNUMA systems, all memory is visible 27 to and accessible from any CPU attached to any cell and cache coherency 28 is handled in hardware by the processor caches and/or the system interconnect. 29 30 Memory access time and effective memory bandwidth varies depending on how far 31 away the cell containing the CPU or IO bus making the memory access is from the 32 cell containing the target memory. For example, access to memory by CPUs 33 attached to the same cell will experience faster access times and higher 34 bandwidths than accesses to memory on other, remote cells. NUMA platforms 35 can have cells at multiple remote distances from any given cell. 36 37 Platform vendors don't build NUMA systems just to make software developers' 38 lives interesting. Rather, this architecture is a means to provide scalable 39 memory bandwidth. However, to achieve scalable memory bandwidth, system and 40 application software must arrange for a large majority of the memory references 41 [cache misses] to be to "local" memory--memory on the same cell, if any--or 42 to the closest cell with memory. 43 44 This leads to the Linux software view of a NUMA system: 45 46 Linux divides the system's hardware resources into multiple software 47 abstractions called "nodes". Linux maps the nodes onto the physical cells 48 of the hardware platform, abstracting away some of the details for some 49 architectures. As with physical cells, software nodes may contain 0 or more 50 CPUs, memory and/or IO buses. And, again, memory accesses to memory on 51 "closer" nodes--nodes that map to closer cells--will generally experience 52 faster access times and higher effective bandwidth than accesses to more 53 remote cells. 54 55 For some architectures, such as x86, Linux will "hide" any node representing a 56 physical cell that has no memory attached, and reassign any CPUs attached to 57 that cell to a node representing a cell that does have memory. Thus, on 58 these architectures, one cannot assume that all CPUs that Linux associates with 59 a given node will see the same local memory access times and bandwidth. 60 61 In addition, for some architectures, again x86 is an example, Linux supports 62 the emulation of additional nodes. For NUMA emulation, linux will carve up 63 the existing nodes--or the system memory for non-NUMA platforms--into multiple 64 nodes. Each emulated node will manage a fraction of the underlying cells' 65 physical memory. NUMA emulation is useful for testing NUMA kernel and 66 application features on non-NUMA platforms, and as a sort of memory resource 67 management mechanism when used together with cpusets. 68 [see Documentation/admin-guide/cgroup-v1/cpusets.rst] 69 70 For each node with memory, Linux constructs an independent memory management 71 subsystem, complete with its own free page lists, in-use page lists, usage 72 statistics and locks to mediate access. In addition, Linux constructs for 73 each memory zone [one or more of DMA, DMA32, NORMAL, HIGH_MEMORY, MOVABLE], 74 an ordered "zonelist". A zonelist specifies the zones/nodes to visit when a 75 selected zone/node cannot satisfy the allocation request. This situation, 76 when a zone has no available memory to satisfy a request, is called 77 "overflow" or "fallback". 78 79 Because some nodes contain multiple zones containing different types of 80 memory, Linux must decide whether to order the zonelists such that allocations 81 fall back to the same zone type on a different node, or to a different zone 82 type on the same node. This is an important consideration because some zones, 83 such as DMA or DMA32, represent relatively scarce resources. Linux chooses 84 a default Node ordered zonelist. This means it tries to fallback to other zones 85 from the same node before using remote nodes which are ordered by NUMA distance. 86 87 By default, Linux will attempt to satisfy memory allocation requests from the 88 node to which the CPU that executes the request is assigned. Specifically, 89 Linux will attempt to allocate from the first node in the appropriate zonelist 90 for the node where the request originates. This is called "local allocation." 91 If the "local" node cannot satisfy the request, the kernel will examine other 92 nodes' zones in the selected zonelist looking for the first zone in the list 93 that can satisfy the request. 94 95 Local allocation will tend to keep subsequent access to the allocated memory 96 "local" to the underlying physical resources and off the system interconnect-- 97 as long as the task on whose behalf the kernel allocated some memory does not 98 later migrate away from that memory. The Linux scheduler is aware of the 99 NUMA topology of the platform--embodied in the "scheduling domains" data 100 structures [see Documentation/scheduler/sched-domains.rst]--and the scheduler 101 attempts to minimize task migration to distant scheduling domains. However, 102 the scheduler does not take a task's NUMA footprint into account directly. 103 Thus, under sufficient imbalance, tasks can migrate between nodes, remote 104 from their initial node and kernel data structures. 105 106 System administrators and application designers can restrict a task's migration 107 to improve NUMA locality using various CPU affinity command line interfaces, 108 such as taskset(1) and numactl(1), and program interfaces such as 109 sched_setaffinity(2). Further, one can modify the kernel's default local 110 allocation behavior using Linux NUMA memory policy. [see 111 Documentation/admin-guide/mm/numa_memory_policy.rst]. 112 113 System administrators can restrict the CPUs and nodes' memories that a non- 114 privileged user can specify in the scheduling or NUMA commands and functions 115 using control groups and CPUsets. [see Documentation/admin-guide/cgroup-v1/cpusets.rst] 116 117 On architectures that do not hide memoryless nodes, Linux will include only 118 zones [nodes] with memory in the zonelists. This means that for a memoryless 119 node the "local memory node"--the node of the first zone in CPU's node's 120 zonelist--will not be the node itself. Rather, it will be the node that the 121 kernel selected as the nearest node with memory when it built the zonelists. 122 So, default, local allocations will succeed with the kernel supplying the 123 closest available memory. This is a consequence of the same mechanism that 124 allows such allocations to fallback to other nearby nodes when a node that 125 does contain memory overflows. 126 127 Some kernel allocations do not want or cannot tolerate this allocation fallback 128 behavior. Rather they want to be sure they get memory from the specified node 129 or get notified that the node has no free memory. This is usually the case when 130 a subsystem allocates per CPU memory resources, for example. 131 132 A typical model for making such an allocation is to obtain the node id of the 133 node to which the "current CPU" is attached using one of the kernel's 134 numa_node_id() or CPU_to_node() functions and then request memory from only 135 the node id returned. When such an allocation fails, the requesting subsystem 136 may revert to its own fallback path. The slab kernel memory allocator is an 137 example of this. Or, the subsystem may choose to disable or not to enable 138 itself on allocation failure. The kernel profiling subsystem is an example of 139 this. 140 141 If the architecture supports--does not hide--memoryless nodes, then CPUs 142 attached to memoryless nodes would always incur the fallback path overhead 143 or some subsystems would fail to initialize if they attempted to allocated 144 memory exclusively from a node without memory. To support such 145 architectures transparently, kernel subsystems can use the numa_mem_id() 146 or cpu_to_mem() function to locate the "local memory node" for the calling or 147 specified CPU. Again, this is the same node from which default, local page 148 allocations will be attempted.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.