On Apr 8, 2011, at 9:02 AM, Hugo Meyer wrote:

Thanks Ralph.

I found a set_lifeline with that i think i solve that error, but, now i'm dealing with another.

[clus3:32001] [[44269,0],2] -> [[44269,1],1] (node: node3) oob-tcp: Number of attempts to create TCP connection has been exceeded.  Can not communicate with peer
Open MPI Error Report:[32001]: While communicating to proc [[44269,1],1] on node node3, proc [[44269,0],2] on node clus3 encountered an error 'Communication failure':OOB Connection retries exceeded.  Can not communicate with peer

I think that this occurs because the daemon [[44269,0],2] doesn't know in wich port and address has been restored the proc. I will look for a way to update this information.

When the proc restarts, it calls orte_routed.init_routes. If you look in routed cm, you should see a call to "register_sync" - this is where the proc sends a message to the local daemon, allowing it to "learn" the port/address where the proc resides.



Best regards.

Hugo

2011/4/6 Ralph Castain <rhc@open-mpi.org>
Looks like the lifeline is still pointing to its old daemon instead of being updated to the new one. Look in orte/mca/routed/cm/routed_cm.c - should be something in there that updates the lifeline during restart of a checkpoint.


On Apr 6, 2011, at 7:50 AM, Hugo Meyer wrote:

Hi all.

I corrected the error with the port. The mistake was because he tried to start theprocess back and the ports are static, the process was taking a port where an app was already running.

Initially, the process was running on [[65478,0],1] and then it moves to [[65478,0],2].

So now i get the socket binded, but i'm getting a communication failure in [[65478,0],1]. I'm sending as an atachment my debug output (there are some things in spanish, but there still are the open-mpi default debug output), where you can see the moment where i kill the process running con clus5 to the moment where it is restored in clus3. And then i get a TERMINATED WITHOUT SYNC in the proc restarted:
clus3:21615] [[65478,0],2] errmgr:orted got state TERMINATED WITHOUT SYNC for proc [[65478,1],1] pid 21705

Here i put the output of my stdout after the socket is binded again when the process restarts.

[1,1]<stdout>:SOCKET BINDED 
[1,1]<stdout>:[clus5:19425] App) notify_response: Waiting for final handshake.
[1,1]<stdout>:[clus5:19425] App) update_status: Update checkpoint status (13, /tmp/radic/1) for [[65478,1],1]
[1,0]<stdout>:INICIEI O BROADCAST (6)
[1,0]<stdout>:FINALIZEI O BROADCAST (6)
[1,0]<stdout>:INICIEI O BROADCAST
[1,3]<stdout>:INICIEI O BROADCAST (6)
[1,2]<stdout>:INICIEI O BROADCAST (6)
[1,3]<stdout>:FINALIZEI O BROADCAST (6)
[1,3]<stdout>:INICIEI O BROADCAST
[1,2]<stdout>:FINALIZEI O BROADCAST (6)
[1,2]<stdout>:INICIEI O BROADCAST
[1,1]<stdout>:[clus5:19425] [[65478,1],1] errmgr:app: job [65478,0] reported state COMMUNICATION FAILURE for proc [[65478,0],1] state COMMUNICATION FAILURE exit_code 1
[1,1]<stdout>:[clus5:19425] [[65478,1],1] routed:cm: Connection to lifeline [[65478,0],1] lost
[1,1]<stdout>:[[65478,1],1] assigned port 31256

Any help on how to solve this error, or how to interpret it will be greatly appreciated.

Best regards.

Hugo

2011/4/5 Hugo Meyer <meyer.hugo@gmail.com>
Hello Ralph and @ll.

Ralph, by following your recomendations i've already restart the process in another node from his checkpoint. But now i'm having a small problem with the oob_tcp. There is the output:

odls_base_default_fns:SETEANDO BLCR CONTEXT
CKPT-FILE: /tmp/radic/1/ompi_blcr_context.13374
ODLS_BASE_DEFAULT_FNS: REINICIO PROCESO EN [[34224,0],2] 
[1,1]<stdout>:INICIEI O BROADCAST (2)
[1,1]<stdout>:[clus5:13374] snapc:single:app do_checkpoint: RESTART (3)
[1,1]<stdout>:[clus5:13374] mca_oob_tcp_init: creating listen socket
[1,1]<stdout>:[clus5:13374] mca_oob_tcp_init: unable to create IPv4 listen socket: Unable to open a TCP socket for out-of-band communications
[1,1]<stdout>:[clus5:13374] App) notify_response: Waiting for final handshake.
[1,1]<stdout>:[clus5:13374] App) update_status: Update checkpoint status (13, /tmp/radic/1) for [[34224,1],1]
[1,0]<stdout>:INICIEI O BROADCAST (6)
[1,0]<stdout>:FINALIZEI O BROADCAST (6)
[1,0]<stdout>:INICIEI O BROADCAST
[1,3]<stdout>:INICIEI O BROADCAST (6)
[1,3]<stdout>:FINALIZEI O BROADCAST (6)
[1,3]<stdout>:INICIEI O BROADCAST
[1,1]<stdout>:[clus5:13374] [[34224,1],1] errmgr:app: job [34224,0] reported state COMMUNICATION FAILURE for proc [[34224,0],1] state COMMUNICATION FAILURE exit_code 1
[1,1]<stdout>:[clus5:13374] [[34224,1],1] routed:cm: Connection to lifeline [[34224,0],1] lost

I'm thinking that this error ocurrs because the process want to create the socket using the port that was previously assigned to it. So, if i want to restart it using another port or something how the other daemons and process will find out about this? Is this a good choice?

Best regards.

Hugo Meyer

2011/3/31 Hugo Meyer <meyer.hugo@gmail.com>
Ok Ralph. 
Thanks a lot, i will resend this message with a new subject.

Best Regards.

Hugo


2011/3/31 Ralph Castain <rhc@open-mpi.org>
Sorry - should have included the devel list when I sent this.


On Mar 30, 2011, at 6:11 PM, Ralph Castain wrote:

I'm not the expert on this area - Josh is, so I'll defer to him. I did take a quick glance at the sstore framework, though, and it looks like there are some params you could set that might help.

"ompi_info --param sstore all"

should tell you what's available. Also, note that Josh created a man page to explain how sstore works. It's in section 7, looks like "man orte_sstore" should get it.


On Mar 30, 2011, at 3:09 PM, Hugo Meyer wrote:

Hello again.

I'm working in the launch code to handle my checkpoints, but i'm a little stuck in how to set the path to my checkpoint and the executable (ompi_blcr_context.PID). I take a look at the code in odls_base_default_fns.c and this piece of code took my attention:

#if OPAL_ENABLE_FT_CR == 1
            /*
             * OPAL CRS components need the opportunity to take action before a process
             * is forked.
             * Needs access to:
             *   - Environment
             *   - Rank/ORTE Name
             *   - Binary to exec
             */
            if( NULL != opal_crs.crs_prelaunch ) {
                if( OPAL_SUCCESS != (rc = opal_crs.crs_prelaunch(child->name->vpid,
                                                                 orte_sstore_base_prelaunch_location,
                                                                 &(app->app),
                                                                 &(app->cwd),
                                                                 &(app->argv),
                                                                 &(app->env) ) ) ) {
                    ORTE_ERROR_LOG(rc);
                    goto CLEANUP;
                }
            }
#endif

But i didn't find out how to set orte_sstore_base_prelaunch_location, i now that initially this is set in the sstore_base_open. For example, as i'm transfering my checkpoint from one node to another, i store the checkpoint that has to be restore in /tmp/1/ and it has a name like ompi_blcr_context.PID.

Is there any function that i didn't see that allows me to do this? I'm asking this because I do not want to change the signature of the functions to pass the details of the checkpoint and the PID.

Best Regards.

Hugo Meyer

2011/3/30 Hugo Meyer <meyer.hugo@gmail.com>
Thanks Ralph.
I have finished the (a) point, and now its working, now i have to work to relaunch from my checkpoint as you said.

Best regards.

Hugo Meyer


2011/3/29 Ralph Castain <rhc@open-mpi.org>
The resilient mapper -only- works on procs being restarted - it cannot map a job for its initial launch. You shouldn't set any rmaps flag and things will work correctly - the default round-robin mapper will map the initial launch, and then the resilient mapper will handle restarts.


On Mar 29, 2011, at 5:18 AM, Hugo Meyer wrote:

Ralph.

I'm having a problem when i try to select the rmaps resilient to be used:

/home/hmeyer/desarrollo/ompi-code/binarios/bin/mpirun -v -np 4 --hostfile ../hostfile --bynode -mca rmaps resilient -mca vprotocol receiver -mca plm rsh -mca routed cm ./coll 6 10 2>out.txt 

I get this as error:
[clus9:25568] [[53334,0],0] hostfile: checking hostfile ../hostfile for nodes
--------------------------------------------------------------------------
Your job failed to map. Either no mapper was available, or none
of the available mappers was able to perform the requested
mapping operation. This can happen if you request a map type
(e.g., loadbalance) and the corresponding mapper was not built.

--------------------------------------------------------------------------
[clus9:25568] errmgr:hnp:update_state() [[53334,0],0]) ------- App. Process state updated for process NULL
[clus9:25568] [[53334,0],0] errmgr:hnp: job [53334,0] reported state NEVER LAUNCHED for proc NULL state UNDEFINED pid 0 exit_code 1
[clus9:25568] [[53334,0],0] errmgr:hnp: job [53334,0] reported state NEVER LAUNCHED
[clus9:25568] [[53334,0],0] errmgr:hnp: abort called on job [53334,0] with status 1

Is there a flag that i'm not turning on? or a component that i should have selected?

Thanks again.

Hugo Meyer


2011/3/26 Hugo Meyer <meyer.hugo@gmail.com>
Ok Ralph.

Thanks a lot for your help, i will do as you said and then let you know how it goes.

Best Regards.

Hugo Meyer


2011/3/25 Ralph Castain <rhc@open-mpi.org>

On Mar 25, 2011, at 10:48 AM, Hugo Meyer wrote:

From what you've described before, I suspect all you'll need to do is add some code in orte/mca/odls/base/odls_base_default_fns.c that (a) checks to see if a process in the launch message is being relocated (the construct_child_list code does that already), and then (b) sends the required info to all local child processes so they can take appropriate action.

Failure detection, re-launch, etc. have all been taken care of for you.


I looked at the code that you mentioned me and i realize that i have two possible options, that i'm going to share with you to know your opinion.

First of all i will let you know my actual situation with the implementation. As i'm working in a Fault Tolerant system, but using uncoordinated checkpoint i'm taking checkpoints of all my process at different time and storing them on the machine where there are residing, but i also send this checkpoints to another node (lets call it protector), so if this node fails his process should be restarted in the protector that have his checkpoints.

Right now i'm detecting the failure of a process and i know where this process should be restarted, and also i have the checkpoint in the protector. And i also have the child information of course.

So, my options are:
First Option

I detect the failure, and then i use orte_errmgr_hnp_base_global_update_state()  with some modifications and the hnp_relocate but changing the spawning to make a restart from a checkpoint, i suposse that using this, the migration of the process to another node will be updated and everyone will know it, because is the hnp who is going to do this (is this ok?).

This is the option I would use. The other one is much, much more work. In this option, you only have to:

(a) modify the mapper so you can specify the location of the proc being restarted. The resilient mapper module will be handling the restart - if you look at orte/mca/rmaps/resilient/rmaps_resilient.c, you can see the code doing the "replacement" and modify accordingly.

(b) add any required info about your checkpoint to the launch message. This gets created in orte/mca/odls/base/odls_base_default_fns.c, the "get_add_procs_data" function (at the top of the file).

(c) modify the launch code to handle your checkpoint, if required - see the file in (b), the "construct_child" and "launch" functions.

HTH
Ralph



Second Option

Modify one of the spawn variations(probably the remote_spawn from rsh) in the PLM framework and then use the orted_comm to command a remote_spawn in the protector, but i don't know here how to update the info so everyone knows about the change or how this is managed.

I might be very wrong in what I said, my apologies if so.

Thanks a lot for all the help.

Best regards.

Hugo Meyer

_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel


_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel


_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel


_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel







<out>