Difference between revisions of "Communicators and Context IDs"

Latest revision as of 16:13, 10 November 2012

What Is A Context ID?

When MPI receives a message and matches it against MPI_Recv requests, it compares the message's envelope to the MPI_Recv's envelope. The envelope is the triple of (source, tag, communicator). The source and tag are explicitly integers, yet the communicator is a logical construct indicating a particular communication context. In MPICH this context is implemented via an additional tag field known as the context id. It's worth remembering that there is no wild card matching for communicators.

The MPICH context ID is a 16-bit integer field that is structured as follows:

VWWWWWWXXXXXLYYZ

In this crude diagram each character represents a bit. There are three fields of the context ID indicated by letter and color:

Dynamic Process Context Bit (V)

If 0, this is a traditional context ID allocated from the context ID mask (explained below). If 1 this is a context ID used for dynamic processes and is allocated in a different manner.

Mask Word Index (W)

This is the index into the context ID mask (explained below).

Bit Index (X)

This is which bit index within the mask word that this ID refers to.

Is Localcomm(L)

This bit is set to 1 if this communicator is a local communicator for an intercommunicator. This allows us to derive local communicator context IDs from the intercommunicator context ID without any communication.

Sub-communicator Type (Y)

This is explained further below.

Context Type Suffix (Z)

This is used to indicate different communication contexts within a communicator. For example, user point-to-point messages (MPI_Send/MPI_Recv) occur in a different context than collective messages (MPI_Bcast, etc). This also explained further below.

The actual type of a context ID is MPIR_Context_id_t, which is typedefed to uint16_t.

All members of a communicator use the same context ID for that communicator, but a context ID is not a globally unique ID. That is, the communicator's group information combined with the context ID constitute a unique ID. A great example of this is when a MPI_Comm_split is called collectively each of the resulting disjoint communicators has the same context ID yet different group information.

Context ID Mask

The context ID mask is a bit vector that is used to keep track of which context IDs have been allocated. In the current code it is an array of MPIR_MAX_CONTEXT_MASK (64) 32-bit unsigned ints for a total of 2048. Each process has its own mask and its state may vary from process to process depending on communicator membership patterns.

Mask Access And Multi-threading

Talk about critical sections, the local context mask, and lowestContextId.

TODO finish this section In the mean time you can examine the code for a better explanation: src/mpi/comm/commutil.c:332

Problems and Gotchas

There are several issues and things to watch out for when working on the context ID code in commutil.c.

• The current code expects that unsigned int values are 32-bits or larger. The comments imply that it needs exactly 32-bit unsigned ints but it looks like we lucked out and it should work with larger sizes as well. This needs to be cleaned up in the current code.
• IDs are allocated from the lowest available mask integer index but the highest available bit index within that integer. This leads to a nice looking pattern when the mask is viewed as a hex string via the MPIR_ContextMaskToStr function but a strange ordering of ID values (124, 120, 116, ..., 0, 252, 248, ..., 128, 380, etc).
• While new IDs are allocated in the fashion described just above, the three default communicators (MPI_COMM_WORLD, MPI_COMM_SELF, and MPIR_ICOMM_WORLD) take up bits 0-2 of word 0 (prefixes 0, 8, and 16). In contrast, the first context ID allocated after MPI_Init will be bit 31 of word 0 (id prefix 124). This works out OK, it's just surprising when you are debugging and get "03fffff8ffffffff..." when you print out the mask field. It wouldn't hurt to change this to something less surprising if we get the time.
• When allocating context IDs for intercommunicators it is important to allocate the recvcontext_id and the context_id fields correctly. By convention the recvcontext_id is allocated from the local group's pool while the context_id is allocated from the remote group's pool. It is important not to reverse this, since MPIR_Free_contextid only frees the recvcontext_id. A reversal of these two fields will result in freeing unallocated IDs and leaking other context IDs in some cases.
• It is safe to call MPIR_Free_contextid on "derived" context IDs (such as in localcomm IDs or topological sub-communicators), however those IDs will not truly be freed at that point. Since they share an entry in the context ID bit vector with the parent context ID they will be ignored. The assumption is that any child IDs will be freed first and then the parent IDs will be freed next, at which point the actual context ID will be released.

Sub-communicator Type Field

This is used to distinguish a "top-level" or "parent" communicator from any "sub-communicators" or "child communicators" it may have. Currently whenever a communicator is created we also create two child communicators for it: an internode comm and an intranode comm. These comms are used to implement SMP-aware collective operations. Rather than calling MPIR_Allreduce three times where we previously called it once, we derive context IDs for the sub-communicators from the parent communicator's context ID. A field value of 0 indicates a parent comm, 1 indicates an intranode communicator, and 2 indicates an internode communicator. See here for the relevant constants: src/include/mpiimpl.h:1368.

Context Type Suffix

The last bit of the ID is used to indicate different communication contexts within a communicator. Point-to-point and collective communication occur in separate contexts and use a different suffix to form different context IDs. Theoretically File or Win communication could use a separate context suffix value, but they don't. Because they don't, we only use 1 bit now instead of 2 as we did previously. The suffix values are named and can be found here: src/include/mpiimpl.h:1283.

These values are passed as the context_offset argument to MPID_Send and friends where it is added to the appropriate context ID stored in the communicator structure. This is the value that is actually used for matching.

The MPIC_* family of functions are convenience functions that put the appropriate collective offset to the corresponding MPID_* function. In practice this always amounts passing "1", but it leaves us the option for more flexible changes in the future.

Context ID API

in src/mpi/comm/commutil.c:0:

static char MPIR_ContextMaskToStr(void)

Useful to dump the state of the context mask.

static void MPIR_Init_contextid(void)

Sets all of the bits of the context mask to 1 except for bits 0,1, and 2 of word 0.

static int MPIR_Locate_context_bit(uint32_t local_mask[])

Finds the highest bit of the lowest word that is set in the given mask and returns the corresponding context ID.

static int MPIR_Allocate_context_bit(uint32_t mask[], MPIR_Context_id_t id)

Clears the bit in mask corresponding to the given context id.

static int MPIR_Find_and_allocate_context_id(uint32_t local_mask[])

Finds the highest bit of the lowest word that is set in the given mask. It resets that bit in the context_mask and returns the found ID prefix.

int MPIR_Get_contextid(MPID_Comm *comm_ptr, MPIR_Context_id_t *context_id)

Allocates a new context ID prefix collectively over the given communicator comm_ptr. Returns the new context ID in context_id. The core of the algorithm copies the current state of the mask to a local buffer and then performs an NMPI_Allreduce with an MPI_BAND operation to find the intersection of valid context IDs among all participating processes. The result of this reduction is fed to MPIR_Find_and_allocate_context_id to determine the new context ID prefix.

int MPIR_Get_intercomm_contextid( MPID_Comm *comm_ptr, MPIR_Context_id_t *context_id, MPIR_Context_id_t *recvcontext_id)

Called by MPIR_Comm_copy to get context IDs for a new intercommunicator from an old intercommunicator. Note that it returns a pair of IDs, one for sending and one for receiving.

When and How Context IDs Are Selected For Communicators

Predefined Communicators

There are three predefined communicators that reserve context IDs at MPI_Init time:

• MPI_COMM_WORLD (id prefix 0)
• MPI_COMM_SELF (id prefix 4)
• MPI_ICOMM_WORLD (id prefix 8)

This occurs here in the code: src/mpi/init/initthread.c:206

Intracommunicators

MPI_Comm_create

Just call MPIR_Get_contextid(old_comm_ptr).

MPI_Comm_split

Call MPIR_Get_contextid(comm_ptr). This ID is the same across all the disjoint communicators that are created. That is, if MPI_Comm_split is called such that three new communicators are created, the context ID will be the same in all three communicators (although the groups will obviously be different between communicators).

MPI_Comm_dup

This calls MPIR_Comm_copy which in turn calls MPIR_Get_contextid over the source communicator. This new context ID is used for the duplicate communicator.

Intercommunicators

MPI_Comm_create

Call MPIR_Get_contextid(old_comm_ptr) and use that as the recvcontext_id. Rank 0s of each group exchange their allocated context IDs and use the received values as the context_id field value.

MPI_Comm_split

Call MPIR_Get_contextid(local_comm_ptr) to get the recvcontext_id for the new communicator. Then roots of the groups exchange context IDs and then broadcast them to the rest of their local groups. This received value serves as the context_id for the new communicator.

All communicators that result from a single collective split call have the same context IDs (but obviously different groups).

MPI_Comm_dup

This calls MPIR_Comm_copy which in turn calls MPIR_Get_intercomm_contextid. Each group generates a recvcontext_id via MPIR_Get_contextid. Then the roots exchange that value with each other and broadcast the result to the local group. The value received from the other side becomes the sending context ID (the field named context_id in the MPID_Comm structure).

MPI_Comm_connect

FIXME WRONG need Pavan to explain the new scheme here.

All connecting processes:

• Allocate a context ID via MPIR_Get_contextid(comm_ptr) over the connecting communicator. This is the recvcontext_id for the new intercommunicator.

Then in the root:

• Connect to the port and create a temporary communicator (context ID 4095 (Why?)) to the root on the other side from this connection.
• Exchange global process group size, local communicator size and the context ID determined locally. This is sent via the temporary communicator.

All:

• broadcast the received info on the local communicator

Just the root:

• exchange PG info with the accept side root

All:

• store the received context ID as the context_id for the new intercommunicator.

Just the root:

• synchronize with the remote root
• free the temporary communicator

All:

• barrier over the local communicator

MPI_Comm_accept

The counterpart to the connect algorithm above. It is essentially the same except the first step is to accept the connection instead of to initiate it.

MPI_Comm_spawn

This is simply implemented via a PMI_Spawn_multiple followed by a MPIR_Comm_connect/MPIR_Comm_accept under the hood.