Ubuntu Manpage: likwid-pin - pin a sequential or threaded application to dedicated processors

NAME

       likwid-pin - pin a sequential or threaded application to dedicated processors

SYNOPSIS

       likwid-pin [-vhSpqim] [-V <verbosity>] [-c/-C <corelist>] [-s <skip_mask>] [-d <delim>]

DESCRIPTION

likwid-pin is a command line application to pin a sequential or multithreaded application to dedicated
processors. It can be used as replacement for taskset. Opposite to taskset no affinity mask but single
processors are specified. For multithreaded applications based on the pthread library the pthread_create
library call is overloaded through LD_PRELOAD and each created thread is pinned to a dedicated processor
as specified in core_list .

Per default every generated thread is pinned to the core in the order of calls to pthread_create it is
possible to skip single threads.

The OpenMP implementations of GCC and ICC compilers are explicitly supported. Clang's OpenMP backend
should also work as it is built on top of Intel's OpenMP runtime library. Others may also work likwid-
pin sets the environment variable OMP_NUM_THREADS for you if not already present. It will set as many
threads as present in the pin expression. Be aware that with pthreads the parent thread is always pinned.
If you create for example 4 threads with pthread_create and do not use the parent process as worker you
still have to provide num_threads+1 processor ids.

likwid-pin supports different numberings for pinning. See section CPU EXPRESSION for details.

For applications where first touch policy on NUMA systems cannot be employed likwid-pin can be used to
turn on interleave memory placement. This can significantly speed up the performance of memory bound
multithreaded codes. All NUMA nodes the user pinned threads to are used for interleaving.

LIKWID introduces the concept of affinity domains which can be described as all hardware threads sharing
a topological entity. There are, for example, affinity domains for each CPU socket in the system and each
of these affinity domains contains the hardware threads that belong to the socket. There is always a
virtual affinity domain 'N' for the whole node, thus all hardware threads. Further names are 'Sx' with
'x' as socket offset ('S0', 'S1', 'S2', ...), 'Dy' with 'y' as CPU die offset, 'Mz' with 'z' as NUMA
domain offset and 'Cx' with 'x' as last level cache offset. It is the offset of the topological entities,
so when there are two sockets with IDs 74 und 10023 (systems like this exist), the socket affinity
domains are 'S0' and 'S1'.

OPTIONS

       -h,--help
              prints a help message to standard output, then exits.

       -v,--version
              prints version information to standard output, then exits.

       -V, --verbose <level>
              verbose output during execution for debugging. 0 for only errors, 1 for  informational  output,  2
              for detailed output and 3 for developer output

       -c,-C <cpu expression>
              specify  a numerical list of processors. The list may contain multiple  items, separated by comma,
              and ranges. For example 0,3,9-11. Other format are available, see the CPU EXPRESSION section.

       -s, --skip <skip_mask>
              Specify skip mask as HEX number. For each set bit the corresponding thread is skipped.

       -S,--sweep
              All ccNUMA memory domains belonging to the specified thread list will be cleaned before  the  run.
              Can solve file buffer cache problems on Linux.

       -p     prints the available thread domains for logical pinning

       -i     set NUMA memory policy to interleave involving all NUMA nodes involved in pinning

       -m     set NUMA memory policy to membind involving all NUMA nodes involved in pinning

       -d <delim>
              usable with -p to specify the CPU delimiter in the cpulist

       -q,--quiet
              silent execution without output

CPU EXPRESSION

1. The most intuitive CPU selection method is a comma-separated list of hardware thread IDs. An example
for this is 0,2 which schedules the threads on hardware threads 0 and 2. The physical numbering also
allows the usage of ranges like 0-2 which results in the list 0,1,2.

2. The CPUs can be selected by their indices inside of an affinity domain. The format is
L:<domain>:<indexlist> for selecting the CPUs inside the given domain. Assuming an virtual affinity
domain 'P' that contains the CPUs 0,4,1,5,2,6,3,7. After sorting it to have physical hardware
threads first we get: 0,1,2,3,4,5,6,7. The logical numbering L:P:0-2 results in the selection 0,1,2
from the physical hardware threads first list.

3. The expression syntax enables the selection according to an selection function with variable input
parameters. The format is either E:<affinity domain>:<numberOfThreads> to use the first
<numberOfThreads> threads in affinity domain <affinity domain> or E:<affinity
domain>:<numberOfThreads>:<chunksize>:<stride> to use <numberOfThreads> threads with <chunksize>
threads selected in row while skipping <stride> threads in affinity domain <affinity domain>.
Examples are E:N:4:1:2 for selecting the first four physical CPUs on a system with 2 hardware thread
per CPU core or E:P:4:2:4 for choosing the first two threads in affinity domain P, skipping 2 threads
and selecting again two threads. The resulting CPU list for virtual affinity domain P is 0,4,2,6

3. The last format schedules the threads not only in a single affinity domain but distributed them
evenly over all available affinity domains of the same kind. In contrast to the other formats, the
selection is done using the physical hardware threads first and then the virtual hardware threads
(aka SMT threads). The format is <affinity domain without number>:scatter like M:scatter to schedule
the threads evenly in all available memory affinity domains. Assuming the two socket domains S0 =
0,4,1,5 and S1 = 2,6,3,7 the expression S:scatter results in the CPU list 0,2,1,3,4,6,5,7 Besides
scatter, there is also 'balanced' (close pinning over domains in natural order) and 'cbalanced'
(close pinning over domain with physical cores first). If we assume the socket setup above, the
expression S:balanced:4 results in the CPU list 0,4,2,6 while with cbalanced, the CPU list is
(assuming 4-7 being SMT threads) 0,1,2,3

EXAMPLE

       1.   For standard pthread application:

       likwid-pin -c 0,2,4-6 ./myApp

       The  parent  process  is  pinned  to  processor 0 which is likely to be thread 0 in ./myApp.  Thread 1 is
       pinned to processor 2, thread 2 to processor 4, thread 3 to processor 5 and thread 4 to processor  6.  If
       more threads are created than specified in the processor list, these threads are pinned to processor 0 as
       fallback.

       2.   For  selection  of  CPUs  inside  of a CPUset only the logical numbering is allowed. Assuming CPUset
            0,4,1,5:

       likwid-pin -c L:N:1,3 ./myApp

       This command pins ./myApp on CPU 4 and the thread started by ./myApp on CPU 5

       3.   A common use-case for the numbering by expression is pinning of an application on the Intel Xeon Phi
            coprocessor with its 60 cores each having 4 SMT threads.

       likwid-pin -c E:N:60:1:4 ./myApp

       This command schedules one applicationn thread per physical CPU core for ./myApp.

IMPORTANT NOTICE

       The detection of shepard threads works for Intel's/LLVM OpenMP runtime (>=12.0), for GCC's OpenMP runtime
       as well as for PGI's OpenMP runtime. If you encounter problems with pinning, please  set  a  proper  skip
       mask  to  skip  the  not-detected shepard threads.  Intel OpenMP runtime 11.0/11.1 requires to set a skip
       mask of 0x1.

AUTHOR

       Written by Thomas Gruber <thomas.roehl@googlemail.com>.

BUGS

       Report Bugs on <https://github.com/RRZE-HPC/likwid/issues>.