On 09/18/2014 04:56 PM, Beichuan Yan wrote:
Rob,
Thank you very much for the suggestion. There are two independent scenarios
using parallel IO in my code:
1. MPI processes conditionally print, i.e., some processes print in current loop (but may not print
in next loop), some processes do not
Thanks for your advice. I added tags for messages in ascending order. But
it didn't work, either.
For example, after 103043 times of communication, in the sender side, it
sends an int 78 with tag 206086, followed by 78 bytes data with tag 206087.
In the receiver side, it receives an int 41 with ta
Rob,
Thank you very much for the suggestion. There are two independent scenarios
using parallel IO in my code:
1. MPI processes conditionally print, i.e., some processes print in current
loop (but may not print in next loop), some processes do not print in current
loop (but may print next loop
There is no guarantee that the messages will be received in the same
order that they were sent.
Use tags or another mechanism to match the messages on send and recv ends.
On 09/18/2014 10:42 AM, XingFENG wrote:
I have found some thing strange.
Basically, in my codes, processes send and receive
On 09/17/2014 05:46 PM, Beichuan Yan wrote:
Hi Rob,
As you pointed out in April that there are many cases that could arouse ADIOI_Set_lock
error. My code writes to a file at a location specified by a shared file pointer (it is a
blocking and collective call): MPI_File_write_ordered(contactFi
If all your send/recv are asynchronous you should ensure different tags to
decipher between messages. It could be that you have the same tag for two
different asynchronous sends, in which case there is a race condition for
the receiving end.
Also, if you know the upper bound of your messages I wou
I have found some thing strange.
Basically, in my codes, processes send and receive messages to/from others
with variable lengths asynchronously. When sending messages, a process
would first send the length of message and then the content of the message.
When receiving, a process would first recei
Yes and no. When I ran a single job that uses 16 MPI processes, and I mapped by
socket and used 8 nodes, 2 ppn, the job ran 30% faster than the same job mapped
by core on 2 nodes. Each process was fairly CPU intensive compared to the
communication, so I suspect that the speed up was due to the f
Lane --
Can you confirm that adding numactl-devel and using --hetero-nodes fixed your
problem?
On Sep 2, 2014, at 5:08 PM, Ralph Castain wrote:
> Argh - yeah, I got confused as things context switched a few too many times.
> The 1.8.2 release should certainly understand that arrangement, an
On Sep 5, 2014, at 11:49 PM, Ralph Castain wrote:
> It would be about the worst thing you can do, to be honest. Reason is that
> each socket is typically a separate NUMA region, and so the shared memory
> system would be sub-optimized in that configuration. It would be much better
> to map-by
Thank you for your reply! I am still working on my codes. I would update
the post when I fix bugs.
On Thu, Sep 18, 2014 at 9:48 PM, Nick Papior Andersen
wrote:
> I just checked, if the tests return "Received" for all messages it will
> not go into memory burst.
> Hence doing MPI_Test will be eno
I just checked, if the tests return "Received" for all messages it will not
go into memory burst.
Hence doing MPI_Test will be enough. :)
Hence, if at any time the mpi-layer is notified about the success of a
send/recv it will clean up. This makes sense :)
See the updated code.
2014-09-18 13:39
ok i have to wait until tomorrow, they have some problems with the
network...
On 09/18/2014 01:27 PM, Nick Papior Andersen wrote:
I am not sure whether test will cover this... You should check it...
I here attach my example script which shows two working cases, and one
not workning (you ca
I am not sure whether test will cover this... You should check it...
I here attach my example script which shows two working cases, and one not
workning (you can check the memory usage simultaneously and see that the
first two works, the last one goes ballistic in memory).
Just check it with tes
Thanks very much for your reply!
To Sir Jeff Squyres:
I think it fails due to truncation errors. I am now logging information of
each send and receive to find out the reason.
To Sir Nick Papior Andersen:
Oh, wait (mpi_wait) is never called in my codes.
What I do is to call MPI_Irecv once. T
In complement to Jeff, I would add that using asynchronous messages
REQUIRES that you wait (mpi_wait) for all messages at some point. Even
though this might not seem obvious it is due to memory allocation "behind
the scenes" which are only de-allocated upon completion through a wait
statement.
20
On Sep 18, 2014, at 2:43 AM, XingFENG wrote:
> a. How to get more information about errors? I got errors like below. This
> says that program exited abnormally in function MPI_Test(). But is there a
> way to know more about the error?
>
> *** An error occurred in MPI_Test
> *** on communicato
By the way, I am using Open MPI 1.6.5 and programming with C++.
On Thu, Sep 18, 2014 at 4:43 PM, XingFENG wrote:
> Dear all,
>
> I am new to MPI. Please forgive me if I ask a redundant question.
>
> I am now programming about graph processing using MPI. I get two problems
> as described below.
>
Dear all,
I am new to MPI. Please forgive me if I ask a redundant question.
I am now programming about graph processing using MPI. I get two problems
as described below.
a. How to get more information about errors? I got errors like below. This
says that program exited abnormally in function *MP
19 matches
Mail list logo
