PTP/planning/remote

From Eclipsepedia

< PTP‎ | planning
Revision as of 08:49, 18 April 2006 by Btibbitts (Talk | contribs)

Jump to: navigation, search

Contents

Evaluation of Related Works for Remote Service Support

List of Authors:

   Tianchao Li (lit@in.tum.de)
   Greg Watson (gwatson@lanl.gov)
   (future authors)

Background

The Need for Remote Compile, Launch and Debug

The current implementation of PTP (v1.0) provides basic support for the coding, compiling, launching and debugging of parallel jobs on a local parallel machine. However, a more typical situation involves one or more of these activities ocurring on a remote machine. In a practical configuration, the development of parallel programs often involves four environments:

User desktop 
the user environment that supports coding and provides UI for compiling, launching and debugging
File server 
the machine on which the project files reside
Build environment 
the machine on which the program is to be built, which provides an environment with the same libraries and services available on the parallel machine
Parallel machine 
the machine on which the program is executed and debugged

It is possible that any combination of these machines may be physically remote, however it is normal for the user's desktop to be remote from the other machines. PTP must be able to support this range of configurations if it is to meet the needs of application developers.

Project Configuration Alternatives

There are three different possible ways that projects may be hosted in Eclipse:

Local Project 
the project and all code of the parallel program is stored in the local file system of the user's desktop.
Remote Project 
the project and all code is physically located on a remote machine (either a dedicated machine or one of the build machine or parallel machine).
Hybrid Project 
the user desktop physically holds the project (in particular the project configuration information) but some or all of the code is physically located on a remote machine.

The process of building, launching and debugging applications differs for each combinations of local and remote resources.

Local Project

In the case where the user's desktop is the central location where the whole project is physically located, the following capabilities are required to support remote build, launch and debug:

Remote Build

  1. The code and build information (e.g. Makefiles) must be first transferred to the build machine (possibly through project export).
  2. The build command is invoked on the build machine.
  3. The build log is transferred back to the local machine and used to update the local models and views (problem view, marker view etc).
  4. The binary is transferred back to the local machine.

Remote Launch & Debug

  1. The control/monitoring/resource management system is invoked on the parallel machine.
  2. The binary is transferred to the parallel machine.
  3. Input data files are transferred to the parallel machine.
  4. The launch command is invoked on the parallel machine.
  5. The user desktop can then control the execution and debugging of the program
  6. The stdio and stderr is trapped and send back to the user desktop.
  7. Any output data files are transferred back to the user desktop.

Remote Project

In the case that a remote machine physically holds the project code, the following capabilities are required to support remote build, launch and debug:

Code Edit

  1. The user environment accesses the project files through remote file system.

Remote Build

  1. The code and build information (e.g. Makefiles) should be transferred to the build machine (possibly through project export), if the build machine is different from the machine that physically holds the project.
  2. The build command is invoked on the build machine.
  3. The build log is transferred back to the local machine and update local models and views (problem view, marker view etc).
  4. The binary is transferred to the remote machine that holds the project.

Remote Launch & Debug

  1. The control/monitoring/resource management system is invoked on the parallel machine.
  2. The binary is transferred to the parallel machine, if different from build machine.
  3. Input data files are transferred to the parallel machine, if different from build machine.
  4. The launch command is invoked on the parallel machine.
  5. The user desktop can then control the execution and debugging of the program
  6. The stdio and stderr is trapped and send back to the user desktop.
  7. Any output data files are transferred back to the build machine, if necessary.

Hybrid Project

If some project files reside on the user's desktop and some on a remote fileserver, then both sets of files need to be reconciled prior to building and/or launching. Once the files have been reconcilde, the hybrid project can be treated as a remote project.

Required Functionality

Additional functionality is required to handle each of the project configuration cases:

Local Project

  1. File transfer between local machine and remote machine (replicte project to remote machine and update project with remote changes, using import/export or team synchronization)
  2. Execute commands on remote machines
  3. Access services on the remote machine (proxy via standard TCP/IP)

Remote Project

  1. Create project on remote file system
  2. Initiate transfer of files between remote machines from the local machine
  3. Execute commands on remote machines
  4. Access services on the remote machine (proxy via standard TCP/IP)

Hybrid Project

  1. Transfer files between local and remote machine and initiate transfer between remote machines
  2. Otherwise, same as remote project

Related Work

There is already some work in both the Eclipse platform and the DSDP project that is relevant to PTP. Almost all of this work is still at a relatively primitive stage however, and we summarize below both their current status and proposed future development targets.

Alternate File System

The ability to support alternative file systems within Eclipse, including non local file systems, has already been requested by a number of people from different areas. Some work has already been undertaken to provide this support.

Status

Starting from Eclipse 3.2 M5, alternative file system support has been partially added. There are two extension points involved:

  • org.eclipse.core.filesystem (defined in org.eclipse.core.filesystem plugin) - core support for file system
  • org.eclipse.ui.ide.filesystemSupport (defined in org.eclipes.ui.ide plugin) - ui support for file system

Existing implementations provided by the Eclipse core for file system include: local, zip, memory. UI supports are also available. The work on FTP file system seems to be going but the current code in the CVS is broken. There is currently no UI extension in the CVS. For more information about the alternate file system, please reference http://www.eclipse.org/eclipse/platform-core/documents/3.2/flexible_workspaces/ and bug item 106176 and 109194.

RSE has also provided an extension to the org.eclipse.core.filesystem extension point, which is a wrapper of its remote file subsystem. However, no extension to org.eclipse.ui.ide.filesystemSupport has been provided and thus it is not possible to use it directly.

Projects and Linked Resources can now be created with alternative file systems (tried with zip file system). The 3.2 M5 release notes say that JDT has support non local file systems. However, no actual support is seen in the distribution.

Future

I expect that the final release of Eclipse 3.2 will provide better support for alternative file systems in both the platform level and the JDT level. However, other projects might need some time to catch up with such a change. Fundamental efforts might even be necessary to provide full support for some of the projects, like CDT, Photran etc.

Remote System Explorer

More specifically, RSE provides UI and API support for browsing and manipulating files on remote systems, execute commands on remote systems, monitoring processes on remote systems and searching

Status

The code currently in the CVS are solely those contributed by IBM. Many of the IDs are still prefixed by "com.ibm.rse". An experiment of the RSE shows that it basically works as it claimed. However, it does not meet our needs and possibly also does not meet the need of DSDP project:

  1. As the name suggests, the RSE is intended to be an explorer (or shell). The whole infrastructure is designed toward this aim. So, instead of trying to integrate and contribute to the underlying Eclipse platform, the RSE sets up a closed environment that are divided into a core and different subsystems. The design of UI components also towards the same target.
  2. The file subsystem and search subsystem of RSE are actually a duplicate functionality with the emerge of alternate file system support from the Eclipse core. Although a wrapper is implemented that exposes RSE file system, its implementation is not complete. Also, no UI integration has been provided.

Besides, there are also other drawbacks in the current implementation of RSE:

  1. The datastore feature is unlikely to be of interest to us, as it requires a specific datastore service to be started on the remote machine.
  2. The support for remote file and remote shell based on SSH is not currently available.
  3. The current code in the CVS combines the core model and UI implementation together in org.eclipse.rse.ui, and thus is difficult to read and understand.

Another possibility to improve RSE is to separate its dependency on the IDE into another plug-in. This will be helpful to use it within RCP programs.

Future

I'm expecting that the file subsystem and search subsystem of RSE should be faded out. The shell subsystem of RSE for remote execution is something also useful for remote file system support in the Eclipse core. So, I'll expect that part be adapted and finally become part of the Eclipse platform. And, only the UI part remains in RSE. RSE is now part of the Target Management subproject of DSDP. Because of the similar challenge for both the PTP and the DSDP, I'm expecting that RSE will be adapted and extended to better support the need of target management for both embedded devices and parallel machines. We will have to collaborate with the developers from DSDP to speed up this process.

Roadmap to PTP Remote Service Support

Based on the current status of Alternative File System and RSE, the following development tasks have been identified:

  1. Extend RSE to support remote execution via SSH [on hold while the status of similar projects is determined]
  2. Provide a remote file system implementation for core.filesystem. Initial implementation will use the 9P2000 protocol.
  3. Extend Eclipse platform to support project export to remote file system.
  4. Extend Eclipse platform to support project update from remote file system.
  5. Make sure the runtime and debug proxy implementations are appropriate for remote invocation.
  6. Modify CDT and Photran project build to support remote build [to be detailed]
  7. Modify PTP launch to support remote launch [to be detailed]
  8. Modify PTP debug to support remote debug [to be detailed]

Summary

RSE can be a potential platform to enable remote service support for PTP. However, huge efforts are still demanded. Collaborative development is also necessary to get the CDT and Photran prepared for remote building and launching.

Open Issues

Resource management is being designed and developed in PTP, how does remote services support fits into the design remains to be discssed.