hwloc-distrib - Man Page

hwloc-distrib [options] <integer>

--single

Singlify each output to a single CPU.

--cpuset-output-format <hwloc|list|taskset> --cof <hwloc|list|taskset>

Change the format of displayed CPU set strings. By default, the hwloc-specific format is used. If list is given, the output is a comma-separated of numbers or ranges, e.g. 2,4-5,8 . If taskset is given, the output is compatible with the taskset program (replaces the former --taskset option).

-v --verbose

Verbose messages.

-i <path>, --input <path>

Read the topology from <path> instead of discovering the topology of the local machine.

If <path> is a file, it may be a XML file exported by a previous hwloc program. If <path> is "-", the standard input may be used as a XML file.

On Linux, <path> may be a directory containing the topology files gathered from another machine topology with hwloc-gather-topology.

On x86, <path> may be a directory containing a cpuid dump gathered with hwloc-gather-cpuid.

When the archivemount program is available, <path> may also be a tarball containing such Linux or x86 topology files.

-i <specification>, --input <specification>

Simulate a fake hierarchy (instead of discovering the topology on the local machine). If <specification> is "node:2 pu:3", the topology will contain two NUMA nodes with 3 processing units in each of them. The <specification> string must end with a number of PUs.

--if <format>, --input-format <format>

Enforce the input in the given format, among xml, fsroot, cpuid and synthetic.

--ignore <type>

Ignore all objects of type <type> in the topology.

--from <type>

Distribute starting from objects of the given type instead of from the top of the topology hierarchy, i.e. ignoring the structure given by objects above.

<type> cannot be among NUMANode, I/O or Misc types.

--to <type>

Distribute down to objects of the given type instead of down to the bottom of the topology hierarchy, i.e. ignoring the structure given by objects below. This may be useful if some latitude is desired for the binding, e.g. just bind several processes to each package without specifying a single core for each of them.

<type> cannot be among NUMANode, I/O or Misc types.

--at <type>

Distribute among objects of the given type.  This is equivalent to specifying both --from and --to at the same time.

--reverse

Distribute by starting with the last objects first, and singlify CPU sets by keeping the last bit (instead of the first bit).

--restrict <cpuset>

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 nodeset=<nodeset>

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.

--restrict-flags <flags>

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).

--disallowed

Include objects disallowed by administrative limitations.

--version

Report version and exit.

-h --help

Display help message and exit.

hwloc-distrib generates a series of CPU masks corresponding to a distribution of a given number of elements over the topology of the machine. The distribution is done recursively from the top of the hierarchy (or from the level specified by option --from) down to the bottom of the hierarchy (or down to the level specified by option --to, or until only one element remains), splitting the number of elements at each encountered hierarchy level not ignored by options --ignore.

This can e.g. be used to distribute a set of processes hierarchically according to the topology of a machine. These masks can be used with hwloc-bind(1).

On hybrid CPUs (or asymmetric platforms), distribution may be suboptimal since the number of cores or PUs inside packages or below caches may vary (the top-down recursive partitioning ignores these numbers until reaching their levels). Hence it is recommended to distribute only inside a single homogeneous domain. For instance on a CPU with energy-efficient E-cores and high-performance P-cores, one should distribute separately N tasks on E-cores and M tasks on P-cores instead of trying to distribute directly M+N tasks on the entire CPUs.

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-distrib's operation is best described through several examples.

If 4 processes have to be distributed across a machine, their CPU masks may be obtained with:

   $ hwloc-distrib 4
   0x0000000f
   0x00000f00
   0x000000f0
   0x0000f000

To distribute only among the second package, the topology should be restricted:

   $ hwloc-distrib --restrict $(hwloc-calc package:1) 4
   0x00000010
   0x00000020
   0x00000040
   0x00000080

To get a single processor of each CPU masks (prevent migration in case of binding)

   $ hwloc-distrib 4 --single
   0x00000001
   0x00000100
   0x00000010
   0x00001000

Each output line may be converted independently with hwloc-calc:

   $ hwloc-distrib 4 --single | hwloc-calc --oo -q -I pu
   PU:0
   PU:8
   PU:4
   PU:12

To convert the output into a list of processors that may be passed to dplace -c inside a mpirun command line:

   $ hwloc-distrib 4 --single | xargs hwloc-calc -I pu
   0,8,4,16

Upon successful execution, hwloc-distrib displays one or more CPU mask strings.  The return value is 0.

hwloc-distrib will return nonzero if any kind of error occurs, such as (but not limited to) failure to parse the command line.

hwloc(7)

hwloc(7), hwloc-bind(1), hwloc-gather-topology(1), hwloc-ps(1).