hwloc-calc - Man Page

hwloc-calc [topology options] [options] <location1> [<location2> [...] ]

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.

--no-smt, --no-smt=<N>

Only keep the first PU per core in the input locations. 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 keeping one of them per core. To rather get PUs #2 to #5 after filtering one per core, you should combine invocations:

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

--cpukind <n>, --cpukind <infoname>=<infovalue>

Only keep PUs whose CPU kind match. Either a single CPU kind is specified as an index, or the info attribute name-value will select matching kinds.

When specified by index, it corresponds to hwloc ranking of CPU kinds which returns energy-efficient cores first, and high-performance power-hungry cores last. The full list of CPU kinds may be seen with lstopo --cpukinds.

Note that this option is applied after searching locations. Hence --cpukind 0 core:1 will return the second core of the machine if it is of kind 0, and nothing otherwise. To rather get the second core among those of kind 0, you should combine invocations:

 hwloc-calc --restrict $(hwloc-calc --cpukind 0 all) core:1

--restrict <cpuset>

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

This is useful when combining invocations to filter some objects before selecting among them.

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.

-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.

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

-p --physical

Use OS/physical indexes instead of logical indexes for both input and output.

-l --logical

Use logical indexes instead of physical/OS indexes for both input and output (default).

--pi --physical-input

Use OS/physical indexes instead of logical indexes for input.

--li --logical-input

Use logical indexes instead of physical/OS indexes for input (default).

--po --physical-output

Use OS/physical indexes instead of logical indexes for output.

--lo --logical-output

Use logical indexes instead of physical/OS indexes for output (default, except for cpusets which are always physical).

-n --nodeset

Interpret both input and output sets as nodesets instead of CPU sets. See --nodeset-output and --nodeset-input below for details.

--no --nodeset-output

Report nodesets instead of CPU sets. This output is more precise than the default CPU set output when memory locality matters because it properly describes CPU-less NUMA nodes, as well as NUMA-nodes that are local to multiple CPUs.

--ni --nodeset-input

Interpret input sets as nodesets instead of CPU sets.

--oo --object-output

When reporting object indexes (e.g. with -I or --local-memory), this option prefixes these indexes with types (e.g. Core:0 instead of 0).

-N --number-of <type|depth>

Report the number of objects of the given type or depth that intersect the CPU set. This is convenient for finding how many cores, NUMA nodes or PUs are available in a machine.

When combined with --nodeset or --nodeset-output, the nodeset is considered instead of the CPU set for finding matching objects. This is useful when reporting the output as a number or set of NUMA nodes.

<type may contain a filter to select specific objects among the type. For instance -N "numa[hbm]" counts NUMA nodes marked with subtype "HBM", while -N "numa[mcdram]" only counts MCDRAM NUMA nodes on KNL.

If an OS device subtype such as gpu  is given instead of osdev, only the os devices of that subtype will be counted.

-I --intersect <type|depth>

Find the list of objects of the given type or depth that intersect the CPU set and report the comma-separated list of their indexes instead of the cpu mask string. This may be used for determining the list of objects above or below the input objects.

When combined with --physical, the list is convenient to pass to external tools such as taskset or numactl --physcpubind or --membind. This is different from --largest since the latter requires that all reported objects are strictly included inside the input objects.

When combined with --nodeset or --nodeset-output, the nodeset is considered instead of the CPU set for finding matching objects. This is useful when reporting the output as a number or set of NUMA nodes.

<type may contain a filter to select specific objects among the type. For instance -N "numa[hbm]" lists NUMA nodes marked with subtype "HBM", while -N "numa[mcdram]" only lists MCDRAM NUMA nodes on KNL.

If an OS device subtype such as gpu is given instead of osdev, only the os devices of that subtype will be returned.

If combined with --object-output, object indexes are prefixed with types (e.g. Core:0 instead of 0).

-H --hierarchical <type1>.<type2>...

Find the list of objects of type <type2> that intersect the CPU set and report the space-separated list of their hierarchical indexes with respect to <type1>, <type2>, etc. For instance, if package.core is given, the output would be Package:1.Core:2 Package:2.Core:3 if the input contains the third core of the second package and the fourth core of the third package.

Only normal CPU-side object types should be used.

NUMA nodes may be used but they may cause redundancy in the output on heterogeneous memory platform. For instance, on a platform with both DRAM and HBM memory on a package, the first core will be considered both as first core of first NUMA node (DRAM) and as first core of second NUMA node (HBM).

--largest

Report (in a human readable format) the list of largest objects which exactly include all input objects (by looking at their CPU sets). None of these output objects intersect each other, and the sum of them is exactly equivalent to the input. No larger object is included in the input.

This is different from --intersect where reported objects may not be strictly included in the input.

--local-memory

Report the list of NUMA nodes that are local to the input objects.

This option is similar to -I numa but the way nodes are selected is different: The selection performed by --local-memory may be precisely configured with --local-memory-flags, while -I numa just selects all nodes that are somehow local to any of the input objects.

If combined with --object-output, object indexes are prefixed with types (e.g. NUMANode:0 instead of 0).

--local-memory-flags

Change the flags used to select local NUMA nodes. Flags may be given as numeric values or as a comma-separated list of flag names that are passed to hwloc_get_local_numanode_objs(). Those names may be substrings of actual flag names as long as a single one matches. The default is 3 (or smaller,larger) which means NUMA nodes are displayed if their locality either contains or is contained in the locality of the given object.

This option enables --local-memory.

--best-memattr <name>

Enable the listing of local memory nodes with --local-memory, but only display the local nodes that have the best value for the memory attribute given by <name> (or as an index).

If the memory attribute values depend on the initiator, the hwloc-calc input objects are 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

If combined with --object-output, the object index is prefixed with its type (e.g. NUMANode:0 instead of 0).

<name> may be suffixed with flags to tune the selection of best nodes, for instance as bandwidth,strict,default. default means that all local nodes are reported if no best could be found. strict means that nodes are selected only if their performance is the best for all the input CPUs. On a dual-socket machine with HBM in each socket, both HBMs are the best for their local socket, but not for the remote socket. Hence both HBM are also considered best for the entire machine by default, but none if strict.

--sep <sep>

Change the field separator in the output. By default, a space is used to separate output objects (for instance when --hierarchical or --largest is given) while a comma is used to separate indexes (for instance when --intersect is given).

--single

Singlify the output to a single CPU.

--cpuset-output-format <hwloc|list|taskset|systemd-dbus-api> --cof <hwloc|list|taskset|systemd-dbus-api>

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). If systemd-dbus-api is given, the output is compatible with systemd's D-Bus API, e.g. "AllowedCPUs ay 0x0002 0x78 0x04" for the CPU set list "3-6,10".

This option has no impact on the format of input CPU set strings, see --cpuset-input-format.

--cpuset-input-format <hwloc|list|taskset> --cif <hwloc|list|taskset>

Change the format of input CPU set strings. By default, the tool tries to guess the type automatically between hwloc, list or taskset formats. This option forces the parsing format to avoid ambiguity for instance when "1,3,5" may be parsed as a hwloc cpuset "0x1,0x00000003,0x00000005" or as list "1-1,3-3,5-5".

This option has no impact on the format of output CPU set strings, see --cpuset-output-format.

-q --quiet

Hide non-fatal error messages. It mostly includes locations pointing to non-existing objects.

-v --verbose

Verbose output.

--version

Report version and exit.

-h --help

Display help message and exit.

hwloc-calc generates and manipulates CPU mask strings or objects. Both input and output may be either objects (with physical or logical indexes), CPU lists (with physical or logical indexes), or CPU mask strings (always physically indexed). Input location specification is described in hwloc(7).

If objects or CPU mask strings are given on the command-line, they are combined and a single output is printed. If no object or CPU mask strings are given on the command-line, the program will read the standard input. It will combine multiple objects or CPU mask strings that are given on the same line of the standard input line with spaces as separators. Different input lines will be processed separately.

Command-line arguments and options are processed in order. First topology configuration options should be given. Then, for instance, changing the type of input indexes with --li or changing the input topology with -i only affects the processing the following arguments.

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-calc utility.

hwloc-calc's operation is best described through several examples.

To display the (physical) CPU mask corresponding to the second package:

   $ hwloc-calc package:1
   0x000000f0

To display the (physical) CPU mask corresponding to the third pacakge, excluding its even numbered logical processors:

   $ hwloc-calc package:2 ~PU:even
   0x00000c00

To display the (physical) CPU mask of the entire topology except the third package:

   $ hwloc-calc all ~package:3
   0x0000f0ff

To combine two (physical) CPU masks:

   $ hwloc-calc 0x0000ffff 0xff000000
   0xff00ffff

To display the list of logical numbers of processors included in the second package:

   $ hwloc-calc --intersect PU package:1
   4,5,6,7

To bind GNU OpenMP threads logically over the whole machine, we need to use physical number output instead:

   $ export GOMP_CPU_AFFINITY=`hwloc-calc --physical-output --intersect PU all`
   $ echo $GOMP_CPU_AFFINITY
   0,4,1,5,2,6,3,7

To display the list of NUMA nodes, by physical indexes, that intersect a given (physical) CPU mask:

   $ hwloc-calc --physical --intersect NUMAnode 0xf0f0f0f0
   0,2

To find how many cores are in the second CPU kind (those cores are likely higher-performance and more power-hungry than cores of the first kind):

   $ hwloc-calc --cpukind 1 -N core all
   4

To convert a cpu mask to human-readable output, the -H option can be used to emit a space-delimited list of locations:

   $ echo 0x000000f0 | hwloc-calc -q -H package.core
   Package:1.Core1 Package:1.Core:1 Package:1.Core:2 Package:1.Core:3

To use some other character (e.g., a comma) instead of spaces in output, use the --sep option:

   $ echo 0x000000f0 | hwloc-calc -q -H package.core --sep ,
   Package:1.Core1,Package:1.Core:1,Package:1.Core:2,Package:1.Core:3

To synthetize a set of cores into largest objects on a 2-node 2-package 2-core machine:

   $ hwloc-calc core:0 --largest
   Core:0
   $ hwloc-calc core:0-1 --largest
   Package:0
   $ hwloc-calc core:4-7 --largest
   L3Cache:1
   $ hwloc-calc core:2-6 --largest
   Package:1 Package:2 Core:6
   $ hwloc-calc pack:2 --largest
   Package:2
   $ hwloc-calc package:2-3 --largest
   L3Cache:1

To get the set of first threads of all cores:

   $ hwloc-calc core:all.pu:0
   0xffff0000
   $ hwloc-calc --no-smt all -I pu
   0,2,4,6,8,10,12,14

To display the list of NUMA nodes, by physical indexes, whose locality is exactly equal to a Package:

   $ hwloc-calc --local-memory-flags 0 --physical-output pack:1
   4,7

To display the best-capacity NUMA node(s), by physical indexes, whose locality is exactly equal to a Package:

   $ hwloc-calc --local-memory-flags 0 --best-memattr capacity --physical-output pack:1
   4

To find the number of NUMA nodes with subtype "HBM":

   $ hwloc-calc -N "numa[hbm]" all
   4

To find the number of NUMA nodes in memory tier 1 (DRAM nodes on a server with HBM and DRAM):

   $ hwloc-calc -N "numa[tier=1]" all
   4

To find the NUMA node of subtype MCDRAM (on KNL) near a PU:

   $ hwloc-calc -I "numa[mcdram]" pu:157
   1

Converting object logical indexes (default) from/to physical/OS indexes may be performed with --intersect combined with either --physical-output (logical to physical conversion) or --physical-input (physical to logical):

   $ hwloc-calc --physical-output PU:2 --intersect PU
   3
   $ hwloc-calc --physical-input PU:3 --intersect PU
   2

One should add --nodeset when converting indexes of memory objects to make sure a single NUMA node index is returned on platforms with heterogeneous memory:

   $ hwloc-calc --nodeset --physical-output node:2 --intersect node
   3
   $ hwloc-calc --nodeset --physical-input node:3 --intersect node
   2

To combine both physical and logical indexes as input:

   $ hwloc-calc PU:2 --physical-input PU:3
   0x0000000c

To display the set of CPUs near network interface eth0:

   $ hwloc-calc os=eth0
   0x00005555

To display the indexes of packages near PCI device whose bus ID is 0000:01:02.0:

   $ hwloc-calc pci=0000:01:02.0 --intersect Package
   1

OS devices may also be filtered by subtype. In this example, there are 8 OS devices in the system, 4 of them are near NUMA node #1, and only 2 of these are CoProcessors:

   $ utils/hwloc/hwloc-calc -I osdev all
   0,1,2,3,4,5,6,7,8
   $ utils/hwloc/hwloc-calc -I osdev node:1
   5,6,7,8
   $ utils/hwloc/hwloc-calc -I coproc node:1
   7,8

To make GNU OpenMP use exactly one thread per core, and in logical core order:

   $ export OMP_NUM_THREADS=`hwloc-calc --number-of core all`
   $ echo $OMP_NUM_THREADS
   4
   $ export GOMP_CPU_AFFINITY=`hwloc-calc --physical-output --intersect PU --no-smt all`
   $ echo $GOMP_CPU_AFFINITY
   0,2,1,3

To export bitmask in a format that is acceptable by the resctrl Linux subsystem (for configuring cache partitioning, etc), apply a sed regexp to the output of hwloc-calc:

   $ hwloc-calc pack:all.core:7-9.pu:0
   0x00000380,,0x00000380   <this format cannot be given to resctrl>
   $ hwloc-calc pack:all.core:7-9.pu:0 | sed -e 's/0x//g' -e 's/,,/,0,/g' -e 's/,,/,0,/g'
   00000380,0,00000380
   # echo 00000380,0,00000380 > /sys/fs/resctrl/test/cpus
   # cat /sys/fs/resctrl/test/cpus
   00000000,00000380,00000000,00000380   <the modified bitmask was corrected parsed by resctrl>

hwloc-calc allows one to generate the very cryptic AllowedCPUs string, which the D-Bus API of systemd expects, from other supported CPU set representations. This is especially useful when the systemd-run command, which understands CPU set provided as list, cannot be used.

First, create a systemd slice:

    $ busctl call org.freedesktop.systemd1 /org/freedesktop/systemd1 org.freedesktop.systemd1.Manager StartUnit ss my_slice.slice fail

Then, configure the CPU set of the slice, using hwloc-calc to translate the syntax:

    $ busctl call org.freedesktop.systemd1 /org/freedesktop/systemd1 org.freedesktop.systemd1.Manager SetUnitProperties 'sba(sv)' my_slice.slice 1 1 $(hwloc-calc pu:0 pu:31 pu:32 pu:63 pu:64 pu:77 --cpuset-output-format systemd-dbus-api)

Finally, add the current process to the slice:

    $ busctl call org.freedesktop.systemd1 /org/freedesktop/systemd1 org.freedesktop.systemd1.Manager StartTransientUnit 'ssa(sv)a(sa(sv))' my_scope.scope fail 3 Delegate b 1 PIDs au 1 $$ Slice s my_slice.slice 0

More info in the org.freedesktop.systemd1(5) manual page.

Upon successful execution, hwloc-calc displays the (physical) CPU mask string, (physical or logical) object list, or (physical or logical) object number list. The return value is 0.

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

hwloc(7), lstopo(1), hwloc-info(1)

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