From HPC Wiki
Jump to: navigation, search


Pinning threads for shared-memory parallelism or binding processes for distributed-memory parallelism is an advanced way to control how your system distributes the threads or processes across the available cores. It is important for improving the performance of your application by avoiding costly remote memory accesses and keeping the threads or processes close to each other. Threads are "pinned" by setting certain OpenMP-related environment variables, which you can do with this command:

$ export <env_variable_name>=<value>

The terms "thread pinning" and "thread affinity" as well as "process binding" and "process affinity" are used interchangeably. You can bind processes by specifying additional options when executing your MPI application.

How to Pin Threads in OpenMP

Schematic of how OMP_PLACES={0}:8:2 would be interpreted
Schematic of how OMP_PROC_BIND=close would be interpreted on a system comprising 2 nodes with 4 hardware threads each
Schematic of OMP_PROC_BIND=spread and an remote memory access of thread 1 accessing the other socket's memory (e.g. thread 0 and thread 1 work on the same data)

OMP_PLACES is employed to specify places on the machine where the threads are put. However, this variable on its own does not determine thread pinning completely, because your system still won't know in what pattern to assign the threads to the given places. Therefore, you also need to set OMP_PROC_BIND.

OMP_PROC_BIND specifies a binding policy which basically sets criteria by which the threads are distributed.

If you want to get a schematic overview of your cluster's hardware, e. g. to figure out how many hardware threads there are, type: $ lstopo.


This variable can hold two kinds of values: a name specifying (hardware) places, or a list that marks places.

Abstract name Meaning
threads a place is a single hardware thread, i. e. the hyperthreading will be ignored
cores a place is a single core with its corresponding amount of hardware threads
sockets a place is a single socket

In order to define specific places by an interval, OMP_PLACES can be set to <lowerbound>:<length>:<stride>. All of these three values are non-negative integers and must not exceed your system's bounds. The value of <lowerbound> can be defined as a list of hardware threads. As an interval, <lowerbound> has this format: {<starting_point>:<length>} that can be a single place, or a place that holds several hardware threads, which is indicated by <length>.

Example hardware OMP_PLACES Places
24 cores with one hardware thread each, starting at core 0 and using every 2nd core {0}:24:2 or {0:1}:24:2 {0}, {2}, {4}, {6}, {8}, {10}, {12}, {14}, {16}, {18}, {20}, {22}
12 cores with two hardware threads each, starting at the first two hardware threads on the first core ({0,1}) and using every 4th core {0,1}:12:4 or {0:2}:12:4 {0,1}, {4,5}, {8,9}, {12,13}, {16,17}, {20,21}

You can also determine these places with a comma-separated list. Say there are 8 cores available with one hardware thread each, and you would like to execute your application on the first four cores, you could define this: $ export OMP_PLACES="{0,1,2,3}"


Now that you have set OMP_PROC_BIND, you can now define the order in which the places should be assigned. This is especially useful for NUMA systems (see references below) because some threads may have to access remote memory, which will slow your application down significantly. If OMP_PROC_BIND is not set, your system will distribute the threads across the nodes and cores randomly.

Value Function
true the threads should not be moved
false the threads can be moved
master worker threads are in the same partition as the master
close worker threads are close to the master in contiguous partitions, e. g. if the master is occupying hardware thread 0, worker 1 will be placed on hw thread 1, worker 2 on hw thread 2 and so on
spread workers are spread across the available places to maximize the space inbetween two neighbouring threads

Options for Binding in Open MPI

Binding processes to certain processors can be done by specifying the options below when executing a program. This is a more advanced way of running an application and also requires knowledge about your system's architecture, e. g. how many cores there are (for an overview of your hardware topology, use $ lstopo). If none of these options are given, default values are set. By overriding default values with the ones specified, you may be able to improve the performance of your application, if your system distributes them in a suboptimal way per default.

Option Function Explanation
--bind-to <arg> bind to the processors associated with hardware component; <arg> can be one of: none, hwthread, core, l1cache, l2cache, l3cache, socket, numa, board; default value: core e. g.: in case of l3cache the processes will be bound to those processors that share the same L3 cache
--map-by <arg> map to the specified hardware component. <arg> can be one of: slot, hwthread, core, L1cache, L2cache, L3cache, socket, numa, board, node, sequential, distance, and ppr; default value: socket if --map-by socket with --bind-to core is used and the program is launched with 4 processes on a two socket machine, process 0 is bound to the first core on socket 0, process 1 is bound to the first core on socket 1, process 2 is bound to the second core on socket 0 and process 3 is bound to the second core on socket 1.
--report-bindings print any bindings for launched processes to the console sample output matching the example for --map-by:
[myhost] MCW rank 0 bound to socket 0[core 0[hwt 0-1]]: [BB/../../../../..][../../../../../..]
[myhost] MCW rank 1 bound to socket 1[core 6[hwt 0-1]]: [../../../../../..][BB/../../../../..]
[myhost] MCW rank 2 bound to socket 0[core 1[hwt 0-1]]: [../BB/../../../..][../../../../../..]
[myhost] MCW rank 3 bound to socket 1[core 7[hwt 0-1]]: [../../../../../..][../BB/../../../..]


Thread affinity in OpenMP

More information on OMP_PLACES and OMP_PROC_BIND

Introduction to OpenMP from PPCES (@RWTH Aachen) Part 3: NUMA & SIMD

FAQ about process affinity in Open MPI