
Project Status Reports
July, 1996
- STATUS OF DOCUMENTATION:
An online demo and online flyer were added to the Web pages. At the
same time, the Adopter's Guide was prepared in its initial draft.
A draft of the User's Guide has also been mounted on the Web.
MQM was included (item 2.3.11) in the
Guidelines for Writing System Software and Tools Requirements
adopted by the National Coordinating Office for HPCC in July, 1996.
A technical report describing MQM is now available through Oregon
State University.
- STATUS OF GRAPHICAL TOOL:
The graphical interface is unchanged since the previous status report.
- INTERFACE BETWEEN IPD AND MQM:
Also unchanged. At the Ptools Annual Meeting, Intel representatives
discussed plans to integrate MQM in future releases of both IPD and XIPD.
- INTERFACE BETWEEN AIMS AND MQM:
Work has begun on the AIMS implementation, as a joint effort involving
NASA/Ames and Los Alamos National Laboratory.
- STANDARD QUERIES:
Unchanged.
December, 1995
- STATUS OF GRAPHICAL TOOL:
A Tk/Tcl graphical interface has been developed in prototype form to show:
- the system overview presenting information of message
queues on all processes at a gross scale
- an overview of message queues on a subset of processes
- a detailed process-level view showing all pending
operations for the process
- a filter window allowing specifying certain messages
to be viewed selectively
- INTERFACE BETWEEN IPD AND MQM:
Integration between Intel's IPD and MQM has been successfully completed.
- STANDARD QUERIES:
Right now, MQM actively queries the message passing system for message
queue information whenever the data is needed. We feel the need to
allow passive querying, which means that MQM queries for message queue
data only when the underlying message passing system or the parallel tool
into which MQM is incorporated allows. This way, the message passing
system or the parallel tool incorporates MQM can be in charge instead of MQM.
June 30, 1994
Active participation by:
- Intel
- NASA/Ames
- Oregon State University
Additional input from:
- PVM group (Al Geist, Jim Kohl, Jack Dongarra)
- MPI group (Rusty Lusk)
- assorted user representatives contacted by Oregon State
The MQM working group was approved in early spring. Work has proceeded
concurrently along several veins, as described below.
- Structure of MQM:
The basic organization of MQM and its interface to the underlying
message system was sketched only vaguely in the proposal. With
input from Intel and NASA/Ames, Oregon State developed a structure that
would isolate the query functions into a separate object layer.
This will allow the modification of queries to suit the messaging
system, without modification of the tool itself. A diagram of
the structure will be found in the WWW pages of the working
group.
- Interface between MQM and XIPD:
Discussions were held between Intel and Oregon State to determine what
additional capability would need to be added to XIPD in order
to interface cleanly to MQM. The isolation of the queries into
a customizable layer will ameliorate most of the problems.
- Interface between MQM and AIMS:
NASA/Ames did preliminary trials to establish the feasibility
of acquiring the needed information on message queues through
the AIMS monitor (AIMS supports either NX or PVM message
passing). All problems appear to be resolvable.
- Interface between MQM and PVM:
Oregon State and the PVM group discussed the feasibility of modifying the
PVM daemon code to handle MQM queries. This would be a simpler
solution for PVM programmers, since they would not need to install
and use AIMS in order to obtain the information. Although PVM
may not support the message deletion feature, they have tentatively
agreed to support the other queries directly.
- Interface between MQM and MPI:
Oregon State and the MPI group discussed the feasibility of doing something
similar with MPI. MPI provides some features to support tool
development; however, their structure would require that all
message operations be traced fully in order to determine the
instantaneous state of the queues. Discussions at the General
Meeting confirmed that this is too high an overhead to incur
for the simple functionality of MQM. MPI will consider future
support for capturing dynamic snapshots of the system, the type
of support needed by MQM.
- Scalability of Visual Representations:
A team at Oregon State addressed the issue of how visual representations
for the user interface would scale for hundreds of processes and
potentially thousands of messages. There is no current tool which
deals with this problem for messages, although a number of other
efforts (chiefly in the area of performance analysis tools) were
examined for potential application to MQM.
The result of the study was the decision to provide hierarchical
display formats: a system overview presenting information on all
processes at a gross scale, and a "zoomed" view presenting detailed
information on a selected process. Both views will support the
filtering operations described in the proposal.
Drafts of the views were presented in the working group's
breakout session at the General Meeting and were assessed as
acceptable.
MQM home page
Parallel Tools at Oregon State home page
Parallel Tools Consortium home page
This document was last updated 13 July '95.
For further information, contact Don Fang,
fangy@research.cs.orst.edu.
July 12, 1995
STATUS OF GRAPHICAL TOOL:
A Tk/Tcl graphical interface has been developed in prototype form to show:
- the system overview presenting
information of message queues on all processes at a gross scale
- an overview of message queues on a subset of processes
- a detailed process-level view showing all pending operations for the process
- a filter window allowing filtering out specific messages
INTERFACE BETWEEN IPD AND MQM:
Integration between Intel's IPD and MQM has been successfully completed.
Nov. 8, 1994
STATUS OF GRAPHICAL TOOL:
A Tk/Tcl graphical interface has been developed in prototype form to show:
- the system overview presenting
information of message queues on all processes at a gross scale
- an overview of message queues on a subset of processes
The primary design questions centered around how to manage the
numerous filtering options in a straightforward, self-explanatory
way. After testing a variety of mechanisms, we decided to divide
filtering options into two types:
- show/hide specific types of operations
- show/hide messages with specific sources/destinations/types
With operation-based filtering options, the user can choose to see
the status of any combination of sends (blocking and/or non-blocking),
messages pending receipt, receives (blocking and/or non-blocking),
and indications of barrier waits. With content-based filtering,
only those messages matching user-specified sources, destinations,
and types will be shown. These filtering mechanisms have been
implemented since the last report.
A query manager that is in charge of querying the message passing
system and managing the message data has also been implemented. MQM
can be hooked up with any message passing systems once their standard
format query is implemented.
Next, we will develop a prototype form of lower-level displays.
That includes view of pending message operations associated with
an individual process, and detailed view of message contents.
The prototype will be demonstrated at the Parallel Tools Consortium
booth at SC'94.
STANDARD QUERY FORMAT:
Since the last report, we've simplified the standard format query.
Sept. 14, 1994
STATUS OF GRAPHICAL TOOL:
A Tk/Tcl graphical interface of the system overview presenting
information on all processes at a gross scale, and an overview of
a subset of processes has been developed in prototype form.
The primary design questions centered around how to manage the
numerous filtering options in a straightforward, self-explanatory
way. A variety of mechanisms were tested, after user input was
employed to develop a user-oriented scheme for grouping messaging
operations into logical subsets.
There was also considerable design work in establishing mechanisms
that would provide intuitive support for process subset filters.
A test version of this was shown to users at NCAR in August, who
concurred that the clearest way was to include an explicit how-to
message directly in the subset window.
One problem with the user interviews is that, since no existing
tool offers "snapshot" views of message queues, users aren't certain
how such a facility would be used in typical programming scenarios.
We hope to have a full prototype ready for user feedback by the end
of September.
STANDARD QUERY FORMAT:
A draft of the standard queries to be supported by the underlying
message system (and the data types required for the queries).
To date, discussion has centered around what, if any, assumptions
might be made about the process/thread identifier format.
Fine-tuning of the query format has not yet begun.
NTERFACE BETWEEN MQM AND AIMS:
An initial draft implementation was made to establish feasibility.
This will work for either NX or PVM message passing, as long as
the application is instrumented by AIMS's instrumenter and linked
with the AIMS monitoring library.
The draft implementation works as follows: Each application sets
up an interrupt handler before commencing execution. On request
by the user (e.g., via ^C), the application is halted by a signal,
triggering activation of a handler which accomplishes the MQM
functional operations:
- delete message that should not have been sent: handler
removes it from someone's message queue
- simulate message (that wasn't really sent): handler
composes a message and sends it
- ignore a receive that should not have been posted: handler
constructs a message that will satisfy the posted receive
- terminate reduction operation: handler simulates the reduction
operation in nodes that failed to participate.
- force completion of a barrier synchronization for which a node
did not check in: handler simulates participation in the barrier.
INTERFACE BETWEEN MQM AND IPD:
Although there has been some discussion about this implementation,
work has not yet begun.