On 05/05/2014 04:57 PM, Kenneth Raffenetti wrote:

Steve,

It looks like the iptables firewall may be active and preventing MPICH from communicating between nodes. You should be able to confirm this by temporarily disabling the firewalls (sudo /etc/init.d/iptables stop) on the involved nodes.

what if steve doesn't have privileges to do so? He can SSH, so there is some ssh tunneling magic he can set up to work around firewalls, right? alas, replace 'tar' with 'ssh tunneling' in this comic and you have me: http://www.xkcd.com/1168/ ==rob

If that's indeed the case, you may want to permanently disable the firewalls or configure them to allow a specific port range for MPICH. The wiki has info on how to tell MPICH which port range to use.

http://wiki.mpich.org/mpich/index.php/Frequently_Asked_Questions#Q:_How_do_I...

Ken

On 05/05/2014 04:39 PM, Rusty Lusk wrote:

...

Hi, can someone help Steve with this problem?

Steve, I am forwarding this to mpich-discuss, which is our problem-reporting list. It will probably go first to the excellent Ken Raffenetti, who I predict will be able to help.

Rusty

Begin forwarded message:

...

*From: *"Steven C. Pieper" <[email protected] <mailto:[email protected]>> *Subject: **Help for installing mpich* *Date: *May 5, 2014 at 4:25:13 PM CDT *To: *Ralph Butler <[email protected] <mailto:[email protected]>>, Rusty Lusk <[email protected] <mailto:[email protected]>> *Reply-To: *<[email protected] <mailto:[email protected]>>

I'm trying to build mpich on the theory linux cluster which is scientific Linux 6.2 which is basically Redhat Enterprise 6

I got mpich-3.1 from the download site, did the configure, make, and install (into ~/mpich-install) . This all seemed OK. I made a machine file:

theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov>:2 # 2 processes on l14 theoryl18.phy.anl.gov <http://theoryl18.phy.anl.gov>:3 # 3 processes on l18

If I am on theoryl14, I can run the hostname test on just l14:

[theoryl14 mpich-test] mpiexec -f machfile_l14 -n 2 hostname theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov> theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov>

Similarly, if I put the l18 line first, I can run 3 processes on l18 from l18.

However I can't run from l14 to l18 or vice versa: Here I start on l14 and specify 3 processes so it has to do one on l18:

[theoryl14 mpich-test] mpiexec -f machfile_l14 -n 3 hostname theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov> theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov> [proxy:0:[email protected]] HYDU_sock_connect (/home/pieper/mpich-3.1/src/pm/hydra/utils/sock/sock.c:172): unable to connect from "theoryl18.phy.anl.gov <http://theoryl18.phy.anl.gov>" to "theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov>" (No route to host) [proxy:0:[email protected]] main (/home/pieper/mpich-3.1/src/pm/hydra/pm/pmiserv/pmip.c:189): unable to connect to server theoryl14.phy.anl.gov <http://theoryl14.phy.anl.gov> at port 48451 (check for firewalls!) ---I hit return here [[email protected]] HYDU_sock_write (/home/pieper/mpich-3.1/src/pm/hydra/utils/sock/sock.c:286): write error (Bad file descriptor) [[email protected]] HYDU_sock_write (/home/pieper/mpich-3.1/src/pm/hydra/utils/sock/sock.c:286): write error (Bad file descriptor) ---I use ctrl-c.

I can ssh without passwords:

[theoryl14 mpich-test] ssh theoryl18.phy.anl.gov <http://theoryl18.phy.anl.gov> hostname theoryl18.phy.anl.gov <http://theoryl18.phy.anl.gov> [theoryl14 mpich-test]

This ssh went through a tunnel that was already established.

So I'm confused. Can one of you help?

BTW: Scientific Linux will allow me to download openmpi but not mpich. I guess I'd like to try to use mpich.

Thanks, Steve

-- Steven C. Pieper:[email protected] Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Morrison, -4100

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

-- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA