RFC0023

Amends and extends:
* RFC0010: Extension of Tool Interaction Support * RFC0001: Provide a mechanism by which tools can interact with a local PMIx server that has opted to accept such connections

Title

Support for Launching Applications Under Debugger Tools

Abstract

Prior RFCs have provided support for several typical use-cases involving tools interacting with applications and the general system management stack. One key use-case, however, has not been covered: the case where an application is actually launched under control of a tool. This RFC defines support for that mode of operation.

Labels

[EXTENSION][CLIENT-API][SERVER-API][RM-INTERFACE][ATTRIBUTES]

Action

[IN PROGRESS]

Copyright (c) 2017-2018 Intel, Inc. All rights reserved.

This document is subject to all provisions relating to code contributions to the PMIx community as defined in the community’s LICENSE file. Code Components extracted from this document must include the License text as described in that file.

Description

Prior RFCs defined mechanisms and standards for connecting debuggers/tools to running jobs, and allowing a tool to itself launch a job, possibly in conjunction with its own daemons. Thus, they dealt with the following use-cases:

Use-case I: attach to an executing application

In this case, the application is started by some other tool – e.g., either an “mpiexec” provided by the programming library, or a native launcher (such as “srun”, “aprun”, or “qsub”) provided by the RM. The debugger is initiated at some random later time, typically when the application is detected as having encountered some problem. An example might look like:

$ mpiexec -n 3 ./myapp &
$ dbgr <pid-of-mpirun>

In this mode of operation, the debugger needs to discover the rendezvous information of the respective launcher and then connect to that PMIx server in the launcher. PMIx already provides an automatic rendezvous protocol for this purpose. Once attached, all PMIx APIs are available for debugger use to obtain information (e.g., the process table of the application), request launch of debugger daemons, etc. – subject to the support provided by the launcher.

Use-case 2: direct-launch an application using a debugger tool

In this case, the dbgr itself can use the PMIx spawn options to control the app’s startup, including directing the RM/app as to when to block and wait for debugger attachment, or stipulating that an interceptor library be preloaded. However, this means that the user is restricted to whatever cmd line options the debugger vendor has provided for operations such as process placement and binding, which places a significant burden on the debugger vendor. An example might look like the following:

$ dbgr -n 3 ./myapp

where the dbgr code would look something like the example provided here. As shown in that example, tool developers are encouraged to query the host RM for its level of support prior to initiating the spawn request. Note that the example takes advantage of the PMIx IO forwarding support to display output from both the application and debugger daemons, and to forward user-typed commands to the debugger daemons for execution.

Assuming it is supported, co-launch of debugger daemons in this use-case is supported by adding a pmix_app_t to the PMIx_Spawn command, indicating that the resulting processes are debugger daemons by setting the PMIX_DEBUGGER_DAEMONS attribute. This tells the RM that the processes are not to be counted against allocated limits, and should not be included in job-level counters (e.g., PMIX_UNIV_SIZE) given to the application (in MPI parlance, the processes spawned by this pmix_app_t are not to be included in MPI_COMM_WORLD). If co-launch is not supported, then the tool can launch the debugger daemons using a separate call to PMIx_Spawn – however, the PMIX_DEBUGGER_DAEMONS attribute still must be provided.

Indirect-launch using a debugger tool

A third use-case that is frequently encountered (but not dealt with by prior RFCs) involves executing a program under a debugger using an intermediate launcher such as mpiexec. In other words, the user executes the application/debugging session using something like the following command line:

$ dbgr mpiexec -n 3 ./myapp

This is an important case for users needing to debug their application from the start of execution, but need/want to use an intermediate launcher. It is a little tricky, however, as it requires some degree of coordination between the dbgr tool and the launcher. Ultimately, it is the launcher that is going to launch the application, and the debugger must somehow inform it (and the application) that this is being done in a debug session so that the application knows to “block” until the debugger attaches to it. There are also potential modifications to the launch directive required by the debugger (e.g., to preload an interceptor library, or to use a debugger-provided local fork/exec agent).

Prior approaches relied on the debugger to modify the launcher’s cmd line, exploiting that launcher’s cmd line options to create the desired behavior. This has been problematic for debugger vendors as every launcher is different, thus requiring the debugger tool to be customized for each implementation. This RFC is focused on eliminating that burden by defining a PMIx-based standard mechanism for the coordination.

Given that the launcher appears on the debugger tool’s cmd line, we assume for the purpose of this use-case that the debugger tool is responsible for launching the launcher itself. This can be done by simply executing the launcher via a command line (e.g., using the Posix “system” function), or the tool may fork/exec it, or may request that it be started by the RM using the PMIx_Spawn API.

Regardless of how it is started, the debugger must set PMIX_LAUNCHER_PAUSE_FOR_TOOL=1 in the environment of mpiexec or in the pmix_info_t array in the spawn command using the PMIX_SET_ENVAR directive. This instructs mpiexec to pause after initialization so it can receive further instructions from the debugger. This might include a request to co-spawn debugger daemons along with the application, or further directives relating to the startup of the application (e.g., to LD_PRELOAD a library, or replace the launcher’s local spawn agent with one provided by the debugger).

As mpiexec starts up, it calls PMIx_server_init to setup its PMIx server. The server initialization includes writing a server-level rendezvous file that allows other processes (such as the originating debugger) to connect to the server. The launcher must then pause, awaiting further instructions from the debugger.

Armed with the pid (returned by fork/exec or the “system” command) or the namespace (returned by PMIx_Spawn) of the executing mpiexec, the debugger tool utilizes the PMIx_tool_connect_server API to complete the connection to the mpiexec server. Note that:

  • PMIx does not allow servers to initiate connections – thus, the debugger tool must initiate the connection to the launcher’s PMIx server.
  • tools can only be connected to one server at a time. Therefore, if connected to the system-level server to use PMIx_Spawn to launch mpiexec, the debugger tool will be disconnected from that server and connected to the PMIx server in mpiexec

At this point, the debugger can execute any PMIx operation, including:

query mpiexec capabilities; pass directives to configure application behavior – e.g., specifying the desired pause point where application processes shall wait for debugger release; request launch of debugger daemons, providing the appropriate pmix_app_t description specify a replacement fork/exec agent; and define/modify standard environmental variables for the application Once ready to launch, mpiexec parses its command line to obtain a description of the desired job. An allocation of resources may or may not have been made in advance (either by the user, or by the tool prior to starting mpiexec)- if not, then mpiexec may itself utilize the PMIx_Alloc API to obtain one from the system-level PMIx server. Once resources are available, mpiexec initiates the launch process by first spawning its daemon network across the allocation – in the above diagram, this is done via ssh commands. After the daemons have launched and wired up, mpiexec sends an application launch command to its daemons, which then start their local client processes and debugger daemons, providing the latter with all information required for them to attach to their targets.

Phase II: Initial tool-launcher connection

Would the debugger agent always be owned by the same user that submitted the job?

What parameters would the debugger agent need to be passed?

What environment variables would the debugger agent need set?

Is it sufficient to place the debugger agent in the same cgroup(s) as the job (without further restricting access to cores)?

When should PBS clean up (kill) debugger agents? When the initial shell for the job exits?

Should the resources used by the debugger agents be accounted for as though they were processes of the job itself?

#define PMIX_SPAWN_LAUNCHER               "pmix.spwn.launcher"          // (bool) app is a user-provided launcher

Modifying Launcher Behavior

There are, in general, two distinct methods by which a tool can start an application under its control:

  • direct launch, whereby the tool calls PMIx_Spawn for the application itself (i.e., not using a launcher such as mpiexec in-between), thereby asking the RM to directly start the processes. In this case, PMIx_Spawn provides adequate controls for specifying the environment of both the application (via the pmix_app_t array) and the launcher itself (via the job_info array).
  • indirect launch, where PMIx_Spawn is given mpiexec as the cmd to execute, which in turn must start the actual application. In this case, the tool does not have direct control over the environment of the application itself as the application appears solely in the argv of mpiexec.

Both cases require the ability to pass instructions to the launcher itself (e.g., replacing fork/exec) that are not covered by envars. This RFC proposes the following behavior-related attributes:

/* Spawn-related Attributes */
#define PMIX_SPAWN_LOCAL_FORK_AGENT       "pmix.spwn.lcl.frk.agnt"      // (char*) path to executable to be used (e.g., in place of fork/exec)

Additional attributes related to the actual launch process (e.g., a method by which the tool can provide directives to the launcher prior to actually starting the application) are described in a companion RFC: Support for Launching Applications Under Debugger Tools

Protoype Implementation

Provide a reference link to the accompanying Pull Request (PR) against the PMIx master repository. If the prototype implementation has been tested against an appropriately modified resource manager and/or client program, then references to those prototypes should be provided. Note that approval of any RFC will be far more likely to happen if such validation has been performed!

Author(s)

Ralph H. Castain
Intel, Inc.
Github: rhc54