152 lines
4.7 KiB
Plaintext
152 lines
4.7 KiB
Plaintext
.\" Man page for mcexec
|
|
.\"
|
|
|
|
.TH MCEXEC 1 "@MCKERNEL_RELEASE_DATE@" "Version @MCKERNEL_VERSION@" MCKERNEL @MCKERNEL_VERSION@"
|
|
.SH NAME
|
|
mcexec \- run an application on McKernel
|
|
.\"
|
|
|
|
.\" ---------------------------- SYNOPSIS ----------------------------
|
|
.SH SYNOPSIS
|
|
.B mcexec \fR [\fIoptions\fR] \fR [\fI<mcos-id>\fR] \fI<command>\fR
|
|
|
|
.\" ---------------------------- DESCRIPTION ----------------------------
|
|
.SH DESCRIPTION
|
|
mcexec executes \fI<command>\fR on McKernel. If <mcos-id> is
|
|
specified, the command is executed on the McKernel instance associated
|
|
with \fI<mcos-id>\fR. Usually, this number is omitted. The options
|
|
are described below.
|
|
|
|
.\" ---------------------------- DESCRIPTION ----------------------------
|
|
.SH OPTIONS
|
|
|
|
.TP
|
|
.B -n N
|
|
Specify the number of MPI ranks run on the node when using MPI (except Fujitsu MPI).
|
|
For example,
|
|
.br
|
|
$ mpirun -n 32 -ppn 4 mcexec -n 4 ./a.out
|
|
.br
|
|
the ./a.out program runs on eight nodes each of
|
|
which has four processes.
|
|
You must specify this option when using MPI (except Fujitsu MPI).
|
|
Undefined behavior entails if not.
|
|
|
|
.TP
|
|
.B -t N
|
|
Specify the number of syscall threads in the proxy process. If this
|
|
option is omitted and the OMP_NUM_THREADS environment is specified,
|
|
the OMP_NUM_THREADS value is used. Otherwise, the number of syscall
|
|
threads equals to the number of CPU cores assigned to McKernel.
|
|
|
|
.\".TP
|
|
.\".B -c N
|
|
.\"Specify the target CPU core on which the application to be run. Do
|
|
.\"not specify this option with -n or with -t.
|
|
|
|
.TP
|
|
.B -h N, --extend-heap-by=N
|
|
Specify the size of heap extension by the brk() system call.
|
|
The default size is 4K Byte. \fIN\fR accepts the following format:
|
|
.br
|
|
.RS 10
|
|
<digits>{k|K|m|M|g|G}
|
|
.br
|
|
e.g.: 10k means 10Kibyte, 100M 100Mibyte, 1G 1Gibyte
|
|
.RE
|
|
.RS 7
|
|
|
|
In certain applications, such as miniFE, specifying "-h 12M" improves
|
|
performance.
|
|
.RE
|
|
|
|
.TP
|
|
.B -s P,[M], --stack-premap=P,[M]
|
|
Pre-map \fIP\fR bytes of stack when creating a process. 2Mibyte is
|
|
used when not specified. And set the max size of stack to \fIM\fR
|
|
bytes. Both of them accept the following format:
|
|
.br
|
|
.RS 10
|
|
<digits>{k|K|m|M|g|G}
|
|
.br
|
|
e.g.: 10k means 10Kibyte, 100M 100Mibyte, 1G 1Gibyte
|
|
.RE
|
|
|
|
.TP
|
|
.B --profile
|
|
Enable system call profiling. After the execution, profiling
|
|
information may be obtained by the ihkosctl tool.
|
|
|
|
.TP
|
|
.B --mpol-no-heap, --mpol-no-stack, --mpol-no-bss,
|
|
Disregard NUMA memory policy in the heap/stack/BSS areas.
|
|
In case of Flat Memory and Quadrant clustering modes on the Xeon Phi
|
|
(KNL), two NUMA nodes, i.e., DDR and MCDRAM, are available.
|
|
The following example in Linux enforces that all memory is allocated
|
|
from NUMA #1, i.e., MCDRAM.
|
|
$ numactl -m 1 ./a.out
|
|
In the following example, memory is allocated in MCDRAM except for the
|
|
stack area. That is, the stack area is allocated in the nearest memory
|
|
of thread/process, i.e., DDR.
|
|
$ mcexec --mpol-no-stack numactl -m 1 ./a.out
|
|
|
|
.TP
|
|
.B --mpol-shm-premap
|
|
Prepopulate shared memory region of the Intel MPI library.
|
|
This option eliminates potential kernel resource contention by
|
|
avoiding page faults in the shared memory region.
|
|
|
|
.TP
|
|
.B -m N, --mpol-threshold=N
|
|
Specify the threshold of memory size for respecting the memory
|
|
allocation policy in NUMA machines. If the size of memory allocation
|
|
is smaller than the one specified in this option, the memory area is
|
|
allocated from the nearest memory of the thread/process disregarding
|
|
the user specified NUMA policy.
|
|
|
|
.TP
|
|
.B --disable-sched-yield
|
|
Disable the sched yield system call. It is useful if the number of
|
|
processes and threads are no more than the number of cores and thus
|
|
explicit scheduling is not required. This option changes the
|
|
LD_PRELOAD environment variable on McKernel, and thus the LD_PRELOAD
|
|
value is not taken over.
|
|
|
|
.TP
|
|
.B --no-bind-ikc-map
|
|
Do not affinitize proxy processes to the IKC target CPUs.
|
|
|
|
The following options will be deprecated.
|
|
.TP
|
|
.B --disable-vdso
|
|
disable the vdso feature. You should not specify this option.
|
|
|
|
.B --enable-vdso
|
|
enable the vdso feature. The default is enable.
|
|
|
|
.PP
|
|
.\" ---------------------------- ENVIRONMENT ----------------------------
|
|
.SH ENVIRONMENT
|
|
.TP
|
|
.B MCKERNEL_RLIMIT_STACK
|
|
If this variable exists, its value is set to the RLIMIT_STACK
|
|
environment variable on Mckernel.
|
|
.TP
|
|
.B MCKERNEL_LD_PRELOAD
|
|
The LD_PRELOAD envirnoment variable in Linux is not taken over by McKernel.
|
|
Instead, if this variable exists, its value is set to the LD_PRELOAD
|
|
environment variable on McKernel. If the --disable-sched-yield option
|
|
is specified in mcexec, LD_PRELOAD defined by the user is not
|
|
effective in this release.
|
|
|
|
.PP
|
|
.\" ---------------------------- SEE ALSO ----------------------------
|
|
.SH SEE ALSO
|
|
\fBihkconfig\fR (1), \fBihkosctl\fR (1)
|
|
|
|
.\" ---------------------------- AUTHORS ----------------------------
|
|
.SH AUTHORS
|
|
Copyright (C) 2017 McKernel Development Team, RIKEN AICS, Japan
|
|
|
|
|