a question about MPI - 未名空间MITBBS历史存档

国际科技财经博客移民网络热点娱乐民生时事公众号

Redian新闻

>未名空间

>Computation - 科学计算

a question about MPI

a question about MPI# Computation - 科学计算

K*N2003-01-11 08:01

1 楼

i have a 3 nodes (each nodes with 32 processors) machine.
but why if i run the following:
*****************************************************
/* Starts MPI processes ... */
MPI_Init(&argc,&argv); /* starts MPI */
MPI_Comm_rank(MPI_COMM_WORLD, &myid); /* get current process id */
MPI_Comm_size(MPI_COMM_WORLD, &p); /* get number of processes */
*****************************************************
p value is always 1?
how to access all 32 processes?
thanks.

h*o2003-01-11 08:01

2 楼

问个问题先...
你怎么跑的这个程序呀...

【在 K*****N 的大作中提到】

: i have a 3 nodes (each nodes with 32 processors) machine.
: but why if i run the following:
: *****************************************************
: /* Starts MPI processes ... */
: MPI_Init(&argc,&argv); /* starts MPI */
: MPI_Comm_rank(MPI_COMM_WORLD, &myid); /* get current process id */
: MPI_Comm_size(MPI_COMM_WORLD, &p); /* get number of processes */
: *****************************************************
: p value is always 1?
: how to access all 32 processes?

K*N2003-01-11 08:01

3 楼

sorry. i am very new to this field. i just downloaded a piece of source
code. and upload to that supercomputer and mpcc my file.
it could pass debug.
i suppose "p" value in that program equals to the number of the processes,
what 's wrong with me? i might have some very silly problem. please let me
know.

【在 h***o 的大作中提到】

: 问个问题先...
: 你怎么跑的这个程序呀...

K*N2003-01-11 08:01

4 楼

got it all. not bother to reply.

【在 K*****N 的大作中提到】

: sorry. i am very new to this field. i just downloaded a piece of source
: code. and upload to that supercomputer and mpcc my file.
: it could pass debug.
: i suppose "p" value in that program equals to the number of the processes,
: what 's wrong with me? i might have some very silly problem. please let me
: know.

h*o2003-01-11 08:01

5 楼

forgot to assign node number in queue script or command line ba
hoho

【在 K*****N 的大作中提到】

: got it all. not bother to reply.

K*N2003-01-11 08:01

6 楼

hehe. no. just much more silly. i didn't use script. hehe.

【在 h***o 的大作中提到】

: forgot to assign node number in queue script or command line ba
: hoho

s*l2003-01-11 08:01

7 楼

Is it possible to detect how much a node is available?
suppose I have to use 10 nodes to do computation, and at
the same time someone else may submit a job to some of the
processors, and I want to send the computations in my program
that cannot be parallelized to a processors that has more
space. For instance I want to have a prog like this:
call mpi_init
...
mpi_scatter
...
mpi_gather
if this processor is fully available
do .....
end if
mpi_finalize

b*g2003-01-11 08:01

8 楼

Sure, use "uptime" command in UNIX. It shows the system load (actually
it is avg length of job queue.)
If you're running on Linux, check the file /proc/loadavg

【在 s*****l 的大作中提到】

: Is it possible to detect how much a node is available?
: suppose I have to use 10 nodes to do computation, and at
: the same time someone else may submit a job to some of the
: processors, and I want to send the computations in my program
: that cannot be parallelized to a processors that has more
: space. For instance I want to have a prog like this:
: call mpi_init
: ...
: mpi_scatter
: ...

s*l2003-01-11 08:01

9 楼

That I know. I meant during the running, in the program unit.

【在 b**g 的大作中提到】

:
: Sure, use "uptime" command in UNIX. It shows the system load (actually
: it is avg length of job queue.)
: If you're running on Linux, check the file /proc/loadavg

l*v2003-01-11 08:01

10 楼

very interesting question. you basically want the slave node doesn't do its
job after the master node has assigning them one.
alternatively, you may want the slave node to check it's load average, respond
to the master node, and let the master decide if to allocate some job to that
node.
you may want to let the master node sort the load average. because you may
have to compete with other programs on all the nodes.
i don't know if it's implementable. if you did that, it would be good to
know.

【在 s*****l 的大作中提到】

: That I know. I meant during the running, in the program unit.