| Dave,
It's not an SDM server generating the routing file, it's the java side of the SDM, then the routing file is transferred to the remote system. We've been debating where this code should live, so I'd welcome input from you on this.
Currently the way it works is like this:
1. User launches debug job. 2. Launch plugin calls debugger initialize() 3. Debugger allocates port number for debugger connection 4. Launch plugin calls submitJob command with debug flag to launch the SDM servers 5. When the SDM processes are all running, the launch plugin calls createDebugSession() 5a. createDebugSession() generates the routing file from the runtime model and transfers it to the remote system 5b. createDebugSession() starts a thread which launches the SDM master process
Now in my mind, steps 5a and 5b are really something that the resource manager should be responsible for, since the proxy may be able to generate the file more efficiently, or may want to start the master in a particular place. However, this would then mean that the proxy will need to know the type of debugger that it is going to launch, since if we're going to be generic we need to support other debuggers, and each debugger may have a different launch sequence.
Anyway, if you have any thoughts or suggestions, they would be appreciated. We can easily change (or disable) the code to work with your proxy.
Regards,
Greg On Aug 26, 2008, at 10:52 PM, Dave Wootton wrote:
Greg
My proxy now starts both the top level SDM as well as the individual process SDMs. I'm actually creating the individual process SDMs (via 'poe sdm ...') before I create the top level SDM since that seemed like a simpler way to fit the creation of these within my existing proxy logic. I'm not sure exactly which gets created first by the system since the individual process SDMs are being invoked in the child process leg following a fork() call in my proxy while the parent leg of the fork() issues a second fork() and invokes the to level SDM on that fork()'s child leg. From what I understand so far about how the SDMs are supposed to do nothing until the routing_file is created, I don't think the timing is a problem.
I have a bug with my code that is supposed to be creating the routing_file that I need to track down.
However, the odd thing going on that I don't understand, is that one of the SDMs is creating the routing file on it's own. I get the following messages that I think are coming from the top level SDM
debug: waiting for connect
PE@k17sf2p03 (RDT): effsize: 3, size: 2, rv: 0
PE@k17sf2p03 (RDT): nodeID: 1, hostname: k17sf2p03, port: 15411
PE@k17sf2p03 (RDT): nodeID: 0, hostname: k17sf2p03, port: 10459
PE@k17sf2p03 (RDT): SDM[2]: [1] No port found for the sdm child. hostname: k17sf2p03
PE@k17sf2p03 (RDT): SDM[1]: sdm_init failed
PE@k17sf2p03 (RDT): 08/26 22:33:54 T(256) Trace: +++ Pid 28627 exited
PACKET:[00000017PE@k17sf2p03 (RDT): 08/26 22:33:58 T(256) Trace: >>> terminate_job entered. (Line 1412)
PE@k17sf2p03 (RDT): 08/26 22:33:58 T(256) Trace: 'jobId=4
PE@k17sf2p03 (RDT): '08/26 22:33:58 T(256) Trace: <<< terminate_job exited. (Line 1438)
PE@k17sf2p03 (RDT): 08/26 22:33:58 T(256) Trace: >>> kill_process entered. (Line 3486)
0000:00000016:00000000]
PE@k17sf2p03 (RDT)ProxyRuntimeClient received event 0 transid=22
calling session finish
debug: received message event
accept thread exiting...
Msg: EventRequestManager - addEventRequest(): Request: Suspend request in status [UNKNOWN] for {0}.
Msg: EventRequestManager - addEventRequest(): Request: Terminate request in status [UNKNOWN] for {0}.
Msg: EventRequestManager - addEventRequest(): Request: Stop debugger request in status [UNKNOWN] for {}.
**** Msg: AbstractEventManager - registerEventRequest(): Request: Suspend request in status [UNKNOWN] for {0}.
**** Msg: AbstractEventManager - notifyEventRequest(): Request: Suspend request in status [ERROR] for {0}.
Error completing debug job launch: Cannot connect to debugger
PACKET:[00000017PE@k17sf2p03 (RDT): 08/26 22:34:04 T(256) Trace: >>> terminate_job entered. (Line 1412)
(RDT is the name for my proxy)
If I look in the current working directory for my proxy, I see a 'routing_file' which has the contents
2
1 k17sf2p03 15411
0 k17sf2p03 10459
This isn't anything my proxy is generating, so I'm not sure where it's coming from.
Any ideas what's going on? Am I misunderstanding what's supposed to be going on?
Dave
Dave,
I didn't write this code, but it sounds like 10 seconds is probably too short for the timeout as you say. It would probably be better to have the master sdm wait forever since it can be killed if the debug launch needs to be aborted.
I believe the race condition should be dealt with already. The first line of the file contains the number of entries, so the SDM will not consider the file complete until it contains this many routing entries. I think it just re-reads the file after some delay until the count is correct.
Greg
On Aug 25, 2008, at 10:46 PM, Dave Wootton wrote:
Greg
I got far enough with my experimentation that I can now get a top level SDM started and not exit. I may have had other problems, but once I turned on sdm debug I found that there's code in the sdm_tcpip_init function that loops for 10 seconds trying to find the routing file (which it looks like is named 'routing_file' in the current directory). If the file isn't found within 10 seconds, sdm issues a timeout message and exits. I changed the timeout to 1000 seconds and the sdm does not exit. So I think I have a starting point to continue working on this.
>From what I understand of the flow you explained, I don't think 10 seconds will be long enough even once I get my proxy to generate the routing_file. As I understand it, I need to create the top SDM, then start the individual task SDMs by 'poe sdm ...', wait for poe to generate the attach.cfg file that gives me the mapping from application task rank to node and pid for each task, then create the routing_file using the attach.cfg file as input, and then the debugger will take off. I think this approach would work for the LoadLeveler case as well, since the attach.cfg file still gets generated. However, on a slow system, or for an application with a large number of tasks, it could take several minutes for this processing to complete.
With a large enough number of tasks, the creation of the routing_file may not complete before the individual SDMs detect it and try to process it. What happens then? Is there logic in the SDM to retry reading the routing_file until it gets a complete copy?
Dave
On Aug 25, 2008, at 11:00 AM, Dave Wootton wrote:
> Greg
> Some additional questions
> 1) It looks like I don't pass the name of the application executable
> as a
> parameter on the top level SDM instance since the top level instance
> isn't
> directly invoking the SDM instances required for individual tasks.
No this isn't necessary. The debugger protocol supplies the executable
name and the application arguments.
>
> 2) What are the invocation parameters of the individual SDM? I'm
> sort of
> guessing I need the hostname and port of the top SDM, the pathname
> of the
> application and any parameters the application requires. I'm
> guessing then
> the individual SDM starts, starts a debugger instance and the debugger
> instance starts the application instance.
The master sdm should be invoked with as 'sdm --host=address --
port=port --debugger=gdb-mi --numprocs=n' where address is the address
of the machine running eclipse and port is a port number assigned by
PTP. The servers will be started with something like 'mpirun sdm -
debugger=gdb-mi --numprocs=n'.
>
> 3) Is the routing file on a node a list of all tasks in the
> application or
> only the tasks running on that node?
A list of all tasks.
>
> 4) How does the routing file get loaded onto each individual node?
At the moment it is assumed there is a shared filesystem. This
requirement will be removed in a later version, and the sdm's
themselves will be used to propagate the routing file.
>
> 5) How does each individual SDM know how to connect back to the top
> SDM if
> the top SDM host/port is not a parameter?
Connections propagate up the tree (starting from the master). Each sdm
knows the index of its children (computed as a binomial tree) so it
just attempts to connect to its children using the address/port
obtained from the routing file.
>
> 6) If the individual SDM is passed the host/port that it connects to
> the
> top SDM, how do I find out what that top level SDM port is?
There is no easy way to do this at the moment, since it is generated
internally and passed to the submitJob command as an argument. The
easiest way would be to print out the arguments to the submitJob
command either in the Java side of the RM or in your proxy.
>
>
> I think I understand how this is supposed to work, and it seems
> reasonable
> for the case where the user specifies a host list file. In the case
> where
> we use LoadLeveler to allocate nodes, I'm not sure how this will work
> since we have no way of knowing what nodes are allocated until the
> poe job
> (the SDMs) starts.
The SDMs do nothing until they get the routing file. Would it be
possible to launch the SDMs, get the node information from LL, then
create the routing file? This is how the new OMPI RM works.
Greg
_______________________________________________
ptp-dev mailing list
ptp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/ptp-dev
_______________________________________________
ptp-dev mailing list
ptp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/ptp-dev
_______________________________________________
ptp-dev mailing list
ptp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/ptp-dev
_______________________________________________
ptp-dev mailing list
ptp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/ptp-dev
|
