1
1
openmpi/ompi/debuggers/MPI_Handles_interface.txt
2015-06-23 20:59:57 -07:00

664 строки
24 KiB
Plaintext

1. George questions the need for the mpidbg_process_t structure (and
the mapping that it provides to MPI_COMM_WORLD). What does Allinea
think of this?
2. 3 Aug 2007: Random notes:
- I kept the prefix MPIDBG/mpidbg (vs. converting to mqs to be the
same as the message queue functionality) to allow this
functionality to be in a different DLL/plugin than the message
queue DLL/plugin.
- I therefore also kept our own MPIDBG return codes (the existing
mqs_* return codes are not sufficient).
3. Some additional open issues throughout the text and header file are
marked with "JMS".
4. 30 Aug 2007: Added questions about uninitialized / already-freed
handles.
***************************************************************************
Premise
=======
Debuggers can display the value of intrinsic datatypes (e.g., int,
double, etc.). Even composite struct instances can be displayed. A
logical extension to this concept is that the debugger should be able
to show information about MPI opaque handles in a running application
(and possibly cache such values for playback when the application is
not running, such as during corefile analysis). Similar in spirit to
the API for obtaining the message passing queues, a simple API can be
used between the debugger and the MPI application to obtain
information about specific MPI handle instances.
***************************************************************************
Background
==========
MPI defines several types of opaque handles that are used in
applications (e.g., MPI_Comm, MPI_Request, etc.). The opaque handles
are references to underlying MPI objects that are private to the MPI
implementation. These objects contain a wealth of information that
can be valuable to display in a debugging context.
Implementations typically have a different underlying type and then
typedef the MPI-specified name to the underlying type. For example,
Open MPI has the following in mpi.h:
typedef struct ompi_communicator_t *MPI_Comm;
The MPI-specified type is "MPI_Comm"; the OMPI-specific type is
"struct ompi_communicator_t *". The debugger cannot simply deduce the
"real" types of MPI handles by looking at the types of well-known
global variables (such as MPI_COMM_WORLD) because these names may be
preprocessor macros in mpi.h; the real name of the symbol may be
implementation-dependent and not easy to guess.
Hence, if "MPI_*" types are unable to be found by the debugger within
the application image, the MPI implementation will need to provide a
list of the actual types to the debugger when the MPI handle
interpretation functionality is initialized. Once the debugger knows
the types of MPI handles, it can provide context-sensitive information
when displaying the values of variables within the application image.
Some MPI implementations use integers for handles in C. Such
implementations are strongly encouraged to use "typedef" for creating
handle types in mpi.h (vs. using #define), such as:
typedef int MPI_Comm;
typedef int MPI_Datatype;
/* etc. */
So that the debugger can identify a variable as an MPI_Comm (vs. an
int) and therefore know that it is an MPI communicator.
In Fortran, however, all MPI handles are defined by the MPI standard
to be of type INTEGER. As such, there is no way for the debugger to
automatically know that a given INTEGER variable is an MPI handle, nor
which kind of MPI handle. It is therefore up to the debugger's UI to
allow users to designate specific INTEGER variables as a given MPI
handle type.
MPI handles can be "freed" by the application, but this actually only
marks the underlying MPI object for freeing; the object itself may not
be freed until all corresponding handles and pending actions have
completed. Additionally, for debugging purposes, an MPI
implementation can choose to *never* free MPI objects (in order to
show that they were marked as freed and/or actually freed).
***************************************************************************
Assumptions
===========
Some terminology:
- host: machine and process on which debugging process is executing.
- debugger: machine and debugging UI, where the debugger is running.
The two machines may be distinct from a hardware point of view,
they may have differing endinanness, word-size, etc.
MPI typically denotes function names in all capitol letters (MPI_INIT)
as a language-neutral form. The Fortran binding for the function is
dependent upon the compiler; the C binding for the function
capitolizes the "MPI" and the first letter of the next token (e.g.,
"MPI_Init"). This text will use the language-neutral names for all
MPI function names.
The debugger will access the handle-interpretation functionality by
loading a plugin provided by the MPI implementation into its process
space. The plugin will contain "query" functions that the debugger
can invoke to obtain information about various MPI handle types. The
query functions generally return the additional information or "not
found" kinds of errors.
The MPI-implementation-provided plugin shares many of the same
characteristics as the Etnus MPI message queue plugin design, but is
loaded slightly differently. The plugin will use the mqs_* functions
defined by the Etnus message queue access interface to read the
process image to obtain MPI object information that is then passed
back to the debugger.
The plugin's query functions should not be called before MPI_INIT has
completed nor after MPI_FINALIZE has started. MPI handles are only
meaningful between the time that MPI_INIT completes and MPI_FINALIZE
starts, anyway.
When MPI handles are marked for freeing by the MPI implementation,
there should be some notice from the debugger that the underlying
objects should not *actually* be freed, but rather orphaned. The
debugger can track these objects and keep a reference count of
how many handles are still referring to the underlying objects. When
the reference count goes to 0, the debugger can call a function in the
application telling the MPI implementation that the object is safe to
be freed.
In this way, valuable debugging information is still available to the
user if they have stale MPI handles because the underlying object will
still be available in memory (marked as "stale"); but MPI object
memory usage is not cumulative.
The debugger may not be able to provide any additional information in
Fortran applications because all MPI handles are of type INTEGER (and
there's no way to tell that any given integer is an MPI communicator
handle, for example) unless the debugger provides some way in the UI
to indicate that a given INTEGER variable is a specific type of MPI
handle.
Note that the following pattern is valid in MPI:
MPI_Request a, b;
MPI_Isend(..., &a);
b = a;
MPI_Request_free(&a);
After executing MPI_REQUEST_FREE, the handle "a" has been set to
MPI_REQUEST_NULL, but the handle "b" still points to the [potentially
ongoing] request. "b" would therefore report that it has been marked
for freeing by the application, but would not report that it was
completed / marked for freeing by the MPI implementation until the
corresponding ISEND actually completes.
The query functions will return newly-allocated structs of information
to the debugger (allocated via mqs_malloc). The debugger will be
responsible for freeing this memory. Arrays/lists of information
contained in the structs must be individually freed if they are not
NULL (i.e., they will each be allocated via mqs_malloc).
Finally, note that not all of this needs to be supported by the
debugger at once. Interpretation of some of the more common handle
types can be implemented first (e.g., communicators and requests),
followed by more types over time.
***************************************************************************
MPI handle types
================
Communicator
============
C: MPI_Comm
C++: MPI::Comm, MPI::Intracomm, MPI::Intercomm, MPI::Cartcomm,
MPI::Graphcomm
A communicator is an ordered set of MPI processes and a unique
communication context. There are 3 predefined communicators
(MPI_COMM_WORLD, MPI_COMM_SELF, MPI_COMM_PARENT), and applications can
create their own communicators. There are two types of communicators:
1. Intracommunicator: contains one group of processes.
Intracommunicators may optionally have a topology:
- Cartesian: a dense, N-dimensional Cartesian topology
- Graph: an arbitrary unweighted, undirected, connected graph
2. Intercommunicator: contains a local group and a remote group of
processes.
--> Information available from communicators:
- String name of length MPI_MAX_OBJECT_NAME
- Flag indicating whether it's a predefined communicator or not
- Whether the handle has been marked for freeing by the
application or not
- This process' unique ID within the communicator ("rank", from 0-(N-1))
- This process' unique ID in the entire MPI universe
- Number of peer processes in the communicator
- Type of communicator: inter or intra
- If Inter, the number and list of peer processes in the communicator
- If intra, whether the communicator has a graph or cartesian
topology
- If have a Cartesian toplogy (mutually exclusive with having a
graph topology), the topology information:
- "ndims", "dims", periods" arguments from the corresponding
call to MPI_CART_CREATE, describing the Cartesian topology.
- If have a graph topology (mutually exclusive with having a
cartesian topology):
- "index" and "edges" arguments from the corresponding call to
MPI_GRAPH_CREATE, describing the nodes and edges in the graph
topology
- Cross reference to the underlying MPI group(s)
- Cross reference to the underlying MPI error handler
- C handle (if available)
- Fortran integer index for this handle (if available)
--> Extra/bonus information that the MPI may be able to provide:
- A list of MPI "attributes" that are attached to the communicator
- A list of any MPI requests currently associated with the
communicator (e.g., ongoing or completed but not released
communicator requests, or, in multi-threaded scenarios, ongoing
communicator operations potentially from other threads)
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- A list of MPI windows using this communicator
- A list of MPI files using this communicator
--> Suggested data types
See mpihandles_interface.h
--> Suggested API functions
See mpihandles_interface.h
---------------------------------------------------------------------------
Datatype
========
C: MPI_Datatype
C++: MPI::Datatype
MPI datatypes are used to express the size and shape of user-defind
messages. For example, a user can define an MPI datatype that
describes a C structure, and can then use that MPI datatype to send
and receive instances (or arrays) of that struct. There are many
predefined datatypes, and applications can create their own datatypes.
--> Information available from datatypes:
- String name
- Flag indicating whether it's a predefined datatype or not
- C handle (if available)
- Fortran integer index for this handle (if available)
- A type map of the overall datatype, composed of *only* intrinsic MPI
datatypes (MPI_INT, MPI_DOUBLE, etc.) that can be rendered by the
debugger into a form similar to MPI-1:3.12
- What function was used to create the datatype:
- <MPI_TYPE_>CREATE_DARRAY, CREATE_F90_COMPLEX,
CREATE_F90_INTEGER, CREATE_F90_REAL, CREATE_HINDEXED,
CREATE_HVACTOR, CREATE_INDEXED_BLOCK, CREATE_RESIZED,
CREATE_STRUCT, CREATE_SUBARRAY
--> JMS: Do we need to differentiate between the MPI-1 and MPI-2
functions? Probably worthwhile, if for nothing other than
completeness (e.g., don't confuse the user saying that a
datatype was created by MPI_TYPE_CREATE_STRUCT when it was
creally created with MPI_TYPE_STRUCT, even though they're
effectively equivalent).
- <MPI_>TYPE_HINDEXED, TYPE_INDEXED, TYPE_HVECTOR, TYPE_VECTOR,
TYPE_STRUCT, TYPE_CONTIGUOUS,
JMS: with the type map provided by MPI, a debugger can show "holes"
in a datatype (potentially indicating missed optimizations by
the application). Very cool/useful!
--> Extra/bonus information that the MPI may be able to provide:
- Ongoing communication actions involving the datatype
(point-to-point, collective, one-sided)
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- Whether the datatype has been "committed" or not
- A list of datatypes used to create this datatype (JMS: may
require caching by the debugger!!)
--> Suggested data types
***TO BE FILLED IN***
--> Suggested API functions
***TO BE FILLED IN***
---------------------------------------------------------------------------
Error handler
=============
C: MPI_Errhandler
C++: MPI::Errhandler
MPI allows applications to define their own error handlers. The
default error handler is to abort the MPI job. Error handlers can be
attached to communicators, files, and windows. There are 3 predefined
error handlers (MPI_ERRORS_ARE_FATAL, MPI_ERRORS_RETURN,
MPI::ERRORS_THROW_EXCEPTIONS), and applications can create their own
error handlers.
--> Information available from error handlers:
- Flag indicating whether it's a predefined error handler or not
- C handle (if available)
- Fortran integer index for this handle (if available)
- Type of errorhandler: communicator, file, window
- If user-defined (i.e., not predefined), the function pointer for
the user function that MPI will invoke upon error
- Whether the callback function is in Fortran or C/C++
--> Extra/bonus information that the MPI may be able to provide:
- String name for predefined handles
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- List of communicators/files/windows that this error handler is
currently attached to
--> Suggested data types
See mpihandles_interface.h
--> Suggested API functions
See mpihandles_interface.h
---------------------------------------------------------------------------
File
====
C: MPI_File
C++: MPI::File
MPI has the concept of parallel IO, where a group of processes
collectively open, read/write, and close files. An MPI_File handle
represents both an ordered set of processes and a file to which they
are accessing. There is one pre-defined file: MPI_FILE_NULL;
applications can open their own files.
--> Information available from files:
- String file name (or "MPI_FILE_NULL")
- Flag indicating whether it's a predefined file or not
- C handle (if available)
- Fortran integer index for this handle (if available)
- Communicator that the file was opened with
- Info key=value pairs that the file was opened with
- Mode that the file was opened with
--> Extra/bonus information that the MPI may be able to provide:
- Whether the handle has been marked for freeing (closing) by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- A list of any MPI requests currently associated with the file
(e.g., ongoing or completed but not released file requests, or, in
multi-threaded scenarios, ongoing file operations potentially from
other threads)
--> Suggested data types
***TO BE FILLED IN***
--> Suggested API functions
***TO BE FILLED IN***
---------------------------------------------------------------------------
Group
=====
C: MPI_Group
C++: MPI::Group
An unordered set of processes. There are predefined and user-defined
groups. Every communicator contains exactly 1 or 2 groups (depending
on the type of communicator). There are 2 predefined groups
(MPI_GROUP_NULL and MPI_GROUP_EMPTY); applications can create their
own groups.
--> Information available from groups:
- C handle (if available)
- Fortran integer index for this handle (if available)
- This process' unique ID in this group
- List of peer processes in this group
--> Extra/bonus information that the MPI may be able to provide:
- String name for predefined handles
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- A list of MPI communicators using this group
--> Suggested data types
***TO BE FILLED IN***
--> Suggested API functions
***TO BE FILLED IN***
---------------------------------------------------------------------------
Info
====
C: MPI_Info
C++: MPI::Info
A set of key=value pairs (the key and value are separate strings) that
can be used to pass "hints" to MPI. There are no predefined info
handles; applications can create their own info handles.
--> Information available from info:
- C handle (if available)
- Fortran integer index for this handle (if available)
- Number of key=value pairs on the info
- List of key=value pairs (each key and value is an individual
string)
--> Extra/bonus information that the MPI may be able to provide:
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- A list of places where the info object is currently being used
--> Suggested data types
***TO BE FILLED IN***
--> Suggested API functions
***TO BE FILLED IN***
---------------------------------------------------------------------------
Request
=======
C: MPI_Request
C++:: MPI::Request, MPI::Grequest, MPI::Prequest
A pointer to an ongoing or completed-but-not-yet-released action.
There are three types of requests:
- Point-to-point communication: non-blocking sends, receives
- File actions: non-blocking reads, writes
- Generalized actions: Users can define their own asynchronous actions
that can be subject to MPI completion semantics
There is one predefined request (MPI_REQUEST_NULL); applications can
create their own requests.
--> Information available from requests:
- Flag indicating whether it's a predefined request or not
- Flag indicating whether the request is persistent or not
- C handle (if available)
- Fortran integer index for this handle (if available)
- Type of the request: pt2pt, file, generalized
- Function that created this request:
- Pt2pt: <MPI_>ISEND, IBSEND, ISSEND, IRSEND, IRECV, SEND_INIT,
BSEND_INIT, SSEND_INIT, RSEND_INIT, RECV_INIT
- File: <MPI_FILE_>IREAD, IREAD_AT, IREAD_SHARED, IWRITE,
IWRITE_AT, IWRITE_SHARED
- Whether the request has been marked "complete" or not
--> Extra/bonus information that the MPI may be able to provide:
- String name for predefined handles
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- Peer process(es) invovled with the request (if available)
- If pt2pt, communicator associated with the request
- If file, file associated with the request
- If pt2pt or file, whether the data transfer has started yet
- If pt2pt or file, whether the data transfer has completed yet
--> Suggested data types
See mpihandles_interface.h
--> Suggested API functions
See mpihandles_interface.h
---------------------------------------------------------------------------
Operation
=========
C: MPI_Op
C++:: MPI::Op
A reduction operator used in MPI collective and one-sided operations
(e.g., sum, multiply, etc.). There are several predefined operators;
applications can also create their own operators.
--> Information available from operators:
- Flag indicating whether it's a predefined operator or not
- C handle (if available)
- Fortran integer index for this handle (if available)
- If user-defined, the function pointer for the user function that
MPI will invoke
- Whether the callback function is in Fortran or C/C++
- Whether the operator is commutative or not
--> Extra/bonus information that the MPI may be able to provide:
- String name for predefined handles
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- List of ongoing collective / one-sided communications associated
with this operator
--> Suggested data types
***TO BE FILLED IN***
--> Suggested API functions
***TO BE FILLED IN***
---------------------------------------------------------------------------
Status
======
C: MPI_Status
C++: MPI::Status
A user-accessible struct that contains information about a completed
communication. The MPI status is a little different from other MPI
handles in that it is the object itselt; not a handle to an underlying
MPI status. For example, if a point-to-point communication was
started with a wildcard receive, the status will contain information
about the peer to whom the communication completed. There are no
predefined statuses.
--> Information available from status:
- Public member MPI_SOURCE: source of the communication
- Public member MPI_TAG: tag of the communication
- Public member MPI_ERROR: error status of the communication
- Number of bytes in the communication
--> Extra/bonus information that the MPI may be able to provide:
- Number of data elements in the communication
--> Suggested data types
See mpihandles_interface.h
--> Suggested API functions
See mpihandles_interface.h
---------------------------------------------------------------------------
Window
======
C: MPI_Win
C++: MPI::Win
An ordered set of processes, each defining their own "window" of
memory for one-sided operations.
--> Information available from windows:
- Communicator that the window was created with
- Base address, length, and displacement units of the window *in this
process*
- Info key=value pairs that the file was opened with
--> Extra/bonus information that the MPI may be able to provide:
- Whether the handle has been marked for freeing by the
application or not
- Whether the underlying object has been "freed" by the MPI
implementation (i.e., made inactive, and would have actually been
freed if not running under a debugger)
- Whether LOCK has been called on this window without a
corresponding UNLOCK yet
- Whether START has been called on this window without a
corresponding COMPLETE yet
- Whether POST has been called on this window without a
corresopnding TEST/WAIT yet
- What the last synchronization call was on the window: FENCE,
LOCK/UNLOCK, START/COMPLETE, POST/TEST/WAIT
--> Suggested data types
***TO BE FILLED IN***
--> Suggested API functions
***TO BE FILLED IN***
---------------------------------------------------------------------------
Address integer
===============
C: MPI_Aint
C++: MPI::Aint
This is an MPI-specific type, but is always an integer value that is
large enough to hold addresses. It is typically an 32 or 64 bits
long. Hence, the debugger should be able to directly display this
value.
---------------------------------------------------------------------------
Offset
======
C: MPI_Offset
C++: MPI::Offset
This is a MPI-specific type, but is always an integer value that is
large enough to hold file offsets. It is typically an 32 or 64 bits
long. Hence, the debugger should be able to directly display this
value.