HWLOC-BIND(1) hwloc HWLOC-BIND(1)

hwloc-bind - Launch a command that is bound to specific processors and/or memory, or consult the binding of an existing program

hwloc-bind [topology options] [options] <location1> [<location2> [...] ] [--] <command> ...

Note that hwloc(7) provides a detailed explanation of the hwloc system and of valid <location> formats; it should be read before reading this man page.

All topology options must be given before all other options.

Only keep the first PU per core before binding. If <N> is specified, keep the <N>-th instead, if any. PUs are ordered by physical index during this filtering.

Note that this option is applied after searching locations. Hence --no-smt pu:2-5 will first select the PUs #2 to #5 in the machine before binding on one of them per core. To rather bind on PUs #2 to #5 after filtering one per core, you should combine with hwloc-calc:


hwloc-bind $(hwloc-calc --restrict $(hwloc-calc --no-smt all) pu:2-5) -- echo hello

Restrict the topology to the given cpuset. This removes some PUs and their now-child-less parents.

Beware that restricting the PUs in a topology may change the logical indexes of many objects, including NUMA nodes.

Restrict the topology to the given nodeset (unless --restrict-flags specifies something different). This removes some NUMA nodes and their now-child-less parents.

Beware that restricting the NUMA nodes in a topology may change the logical indexes of many objects, including PUs.

Enforce flags when restricting the topology. Flags may be given as numeric values or as a comma-separated list of flag names that are passed to hwloc_topology_restrict(). Those names may be substrings of actual flag names as long as a single one matches, for instance bynodeset,memless. The default is 0 (or none).
Include objects disallowed by administrative limitations.
Select the best NUMA node among the given memory binding set by looking at the memory attribute given by <name> (or as an index).

If the memory attribute values depend on the initiator, the CPU binding set is used as the initiator.

Standard attribute names are Capacity, Locality, Bandwidth, and Latency. All existing attributes in the current topology may be listed with


$ lstopo --memattrs

Only take high bandwidth memory nodes (marked with "HBM" subtype, or "MCDRAM" on Intel Xeon Phi) in account when looking for NUMA nodes in the input locations.

This option must be combined with NUMA node locations, such as --hbm numa:1 for binding on the second HBM node. It may also be written as numa[hbm]:1 or numa[mcdram]:1.

Ignore high bandwidth memory nodes (marked with "HBM" subtype, or "MCDRAM" on Intel Xeon Phi MCDRAM) when looking for NUMA nodes in the input locations.

All these options must be given after all topology options above.

Use following arguments for CPU binding (default).
Use following arguments for memory binding. If --mempolicy is not also given, the default policy is bind.
Change the memory binding policy.

This option is only meaningful when an actual binding is also given with --membind. If --membind is given without --mempolicy, the default policy is bind.

The available policies are default, firsttouch, bind, interleave and nexttouch. See hwloc.h for details about these policies.

Note that hwloc's memory binding policies may be slightly different from operating system policies. For instance, the hwloc bind policy uses Linux MPOL_PREFERRED_MANY (or MPOL_PREFERRED) by default, but it switches to Linux MPOL_BIND if the hwloc strict option or flag is also given.

Report the current bindings. The output is an opaque bitmask that may be translated into objects with hwloc-calc (see EXAMPLES below).
 
When a command is given, the binding is displayed before executing the command. When no command is given, the program exits after displaying the current binding.
 
When combined with --membind, report the memory binding instead of CPU binding.
 
No location may be given since no binding is performed.

Report binding as a NUMA memory node set instead of a CPU set if --get was given. This is useful for manipulating CPU-less NUMA nodes since their cpuset is empty while their nodeset is correct.
 
Also parse input bitmasks as nodesets instead of cpusets.
 
When this option is not passed, individual input bitmasks may still be parsed as nodesets if they are prefixed with nodeset=.

Report the last processors where the process ran. The output is an opaque bitmask that may be translated into objects with hwloc-calc (see EXAMPLES below).
 
Note that the result may already be outdated when reported since the operating system may move the process to other processors at any time according to the binding.
 
When a command is given, the last processors is displayed before executing the command. When no command is given, the program exits after displaying the last processors.
 
This option cannot be combined with --membind.
 
No location may be given since no binding is performed.

Bind on a single CPU to prevent migration.
Require strict binding.
Operate on pid <pid>
Operate on thread <tid> instead of on an entire process. The feature is only supported on Linux for thread CPU binding, or for reporting the last processor where the thread ran if -e was also passed.
Interpret input locations with OS/physical indexes instead of logical indexes. This option does not apply to the output, see --get above.
Interpret input locations with logical indexes instead of physical/OS indexes (default). This option does not apply to the output, see --get above.
Display CPU set strings in the format recognized by the taskset command-line program instead of hwloc-specific CPU set string format. This option has no impact on the format of input CPU set strings, both formats are always accepted.
Launch the executable even if binding failed.
Hide non-fatal error messages. It includes locations pointing to non-existing objects, as well as failure to bind. This is usually useful in addition to --force.
Verbose output.
Report version and exit.
Display help message and exit.

hwloc-bind execs an executable (with optional command line arguments) that is bound to the specified location (or list of locations). Location specification is described in hwloc(7). Upon successful execution, hwloc-bind simply sets bindings and then execs the executable over itself.

If a bitmask location is given with prefix nodeset=, then it is considered a nodeset instead of a CPU set. See also --nodeset.

If multiple locations are given, they are combined in the sense that the binding will be wider. The process will be allowed to run on every location inside the combination.

The list of input locations may be explicitly ended with "--".

If binding fails, or if the binding set is empty, and --force was not given, hwloc-bind returns with an error instead of launching the executable.

NOTE: It is highly recommended that you read the hwloc(7) overview page before reading this man page. Most of the concepts described in hwloc(7) directly apply to the hwloc-bind utility.

hwloc-bind's operation is best described through several examples. More details about how locations are specified on the hwloc-bind command line are described in hwloc(7).

To run the echo command on the first logical processor of the second package:


$ hwloc-bind package:1.pu:0 -- echo hello

which is exactly equivalent to the following line as long as there is no ambiguity between hwloc-bind option names and the executed command name:


$ hwloc-bind package:1.pu:0 echo hello

To bind the "echo" command to the first core of the second package and the second core of the first package:


$ hwloc-bind package:1.core:0 package:0.core:1 -- echo hello

To bind on the first PU of all cores of the first package:


$ hwloc-bind package:0.core:all.pu:0 -- echo hello
$ hwloc-bind --no-smt package:0 -- echo hello

To bind on the memory node local to a PU with largest capacity:


$ hwloc-bind --best-memattr capacity --cpubind pu:23 --membind pu:23 -- echo hello

To bind memory on the first NUMA node marked with "HBM" subtype:


$ hwloc-bind --membind numa[hbm]:0 -- echo hello
$ hwloc-bind --hbm --membind numa:0 -- echo hello

To bind memory on the first high-bandwidth memory node (MCDRAM) on Intel Xeon Phi:


$ hwloc-bind --membind numa[mcdram]:0 -- echo hello
$ hwloc-bind --hbm --membind numa:0 -- echo hello

Note that binding the "echo" command to multiple processors is probably meaningless (because "echo" is likely implemented as a single-threaded application); these examples just serve to show what hwloc-bind can do.

To run on the first three packages on the second and third nodes:


$ hwloc-bind node:1-2.package:0:3 -- echo hello

which is also equivalent to:


$ hwloc-bind node:1-2.package:0-2 -- echo hello

Note that if you attempt to bind to objects that do not exist, hwloc-bind will not warn unless -v was specified.

To run on processor with physical index 2 in package with physical index 1:


$ hwloc-bind --physical package:1.core:2 -- echo hello

To run on odd cores within even packages:


$ hwloc-bind package:even.core:odd -- echo hello

To run on the first package, except on its second and fifth cores:


$ hwloc-bind package:0 ~package:0.core:1 ~package:0.core:4 -- echo hello

To run anywhere except on the first package:


$ hwloc-bind all ~package:0 -- echo hello

To run on a core near the network interface named eth0:


$ hwloc-bind os=eth0 -- echo hello

To run on a core near the PCI device whose bus ID is 0000:01:02.0:


$ hwloc-bind pci=0000:01:02.0 -- echo hello

To bind memory on second memory node and run on first node (when supported by the OS):


$ hwloc-bind --cpubind node:1 --membind node:0 -- echo hello

hwloc-bind does not have an option to select a kind of CPU core but it may be combined with hwloc-calc to do so. For instance, to bind on the first two cores whose kind matches CoreType=IntelAtom:


$ hwloc-bind $(hwloc-calc --restrict $(hwloc-calc --cpukind CoreType=IntelAtom all) core:0-1) -- echo hello

The --get option can report current bindings. This example shows nesting hwloc-bind invocations to set a binding and then report it:


$ hwloc-bind node:1.package:2 -- hwloc-bind --get
0x00004444,0x44000000

hwloc-calc can also be used to convert cpu mask strings to human-readable package/core/PU strings; see the description of -H in hwloc-calc(1) for more details. The following example binds to all the PUs in a specific core, uses the --get option to retrieve where the process was actually bound, and then uses hwloc-calc to display the resulting cpu mask in space-delimited list of human-readable locations:


$ hwloc-bind package:1.core:2 -- hwloc-bind --get | hwloc-calc -H package.core.pu
Package:1.Core:2.PU:0 Package:1.Core:2.PU:1

hwloc-calc may convert this output into actual objects, either with logical or physical indexes:


$ hwloc-calc --physical -I pu `hwloc-bind --get`
26,30,34,38,42,46
$ hwloc-calc --logical -I pu `hwloc-bind --get` --sep " "
24 25 26 27 28 29

Locations may also be specified as a hex bit mask (typically generated by hwloc-calc). For example:


$ hwloc-bind 0x00004444,0x44000000 -- echo hello
$ hwloc-bind `hwloc-calc node:1.package:2` -- echo hello

The current memory binding may also be reported:


$ hwloc-bind --membind node:1 --mempolicy interleave -- hwloc-bind --get --membind
0x000000f0 (interleave)

If the graphics-enabled lstopo is available, use for instance


$ hwloc-bind core:2 -- lstopo --pid 0

to check what the result of your binding command actually is. lstopo will graphically show where it is bound to by hwloc-bind.

Upon successful execution, hwloc-bind execs the command over itself. The return value is therefore whatever the return value of the command is.

hwloc-bind will return nonzero if any kind of error occurs, such as (but not limited to): failure to parse the command line, failure to retrieve process bindings, or lack of a command to execute.

hwloc(7), lstopo(1), hwloc-calc(1), hwloc-distrib(1)

December 4, 2023 2.10.0