Signaling Communication Events in a Computer Network

The present invention is related to communications in a computer network and is more particularly related to signalling various events associated with non-blocking communications calls initiated by a task of a parallel job in a distributed parallel data processing system.

In a distributed parallel data processing system a communication mechanism is provided to allow users to communicate between different tasks of a parallel job. Typically to alleviate the effect of communication delays a user is allowed to overlap computation with communication by the use of nonblocking communication calls. When a nonblocking call is invoked, the user needs to know when the call in question has "completed". There are many reasons to know that a nonblocking call has completed--for instance the user may have to free up resources used in the non-blocking calls or the user may want to reuse buffers provided to the nonblocking call.

In order to permit a high degree of concurrency, a low-level application programming interface (LAPI), to be described, defines three types of events associated with nonblocking calls which indicate whether (1) resources of a message are available at origin (2) resources of a message are available at target and (3) a particular message has completed (as related to the origin and target) with respect to the requestor.

Various solutions have been proposed to recognize events associated with nonblocking communication calls. Most previous solutions (MPI is one example) require a distinct handle per message of the user. A distinct handle per message requires more storage and may also be less efficient because it requires the user to check on each of these individually. MPI attempts to alleviate this problem by providing "wait.sub.-- any" and "wait.sub.-- all". However these still require a handle per message and thus the implementation is also expected to be more complex.

Further, MPI has a notion of standard and synchronous messages in order to allow the user to specify the type of nonblocking behavior required. This makes the use of the API more cumbersome since it burdens the user to have to select a particular interface.

The present invention provides a convenient and efficient mechanism for signalling to the user various events associated with nonblocking communication calls initiated by a task of a parallel job thus allowing the user a common quick method to check on communication events.

The present invention uses counters to signal the events associated with the nonblocking communication calls. The specified counter is incremented whenever the event associated with the counter occurs. By checking the value of the counter the user can determine if the event (associated with the counter by the user) in question has occurred.

In the present invention, interfaces that involve nonblocking communication take in as parameters different counters to signal events associated with the communication call in question. The counter structure is defined, and the user is expected to access the counter only using the defined interfaces. However, the user specifies the counters and the relationship of a counter to a message/event. The user may use the same counter across multiple messages. This gives the user the freedom to group different communication calls with the same counter and check their completion as a group.

A library which updates the user specified counters when a particular event (or one of the events) with which the counter was associated has occurred. The user can check on the status of a particular event by polling on the value of the appropriate counter; this polling is done using a LAPI specified interface. If a user wants to wait on a particular event rather than check if it has occurred, LAPI provides an interface to wait on the counter until the counter reaches the specified value. This is a blocking interface. If the user waits for the counter to reach a specific value, then when LAPI returns from the wait, the counter value is decremented by the value (which indicates the number of events) specified in the wait.

The present invention allows the same counter to be associated with multiple messages.

By providing the counter solution, a single interface representing the MPI standard and synchronous calls are provided. However, the ability to specify different types of nonblocking behavior is still maintained. The user can simulate either standard or synchronous behavior by specifying appropriate counters and waiting on them at the appropriate places in the code. And the user can mix the use as needed by the application making this more flexible.

The present invention provides flexibility to the user that is not provided in the other solutions. Further the present invention enables efficient implementations and uses less storage.

The foregoing and other objects, features and advantages of the invention will be apparent from the following more particular description of the preferred embodiment of the invention as illustrated in the drawings.

FIG. 1 is a block diagram of a distributed, parallel data processing system usable with the present invention;

FIG. 2 is a block diagram showing the LAPI layer of the present invention in the operating system of each node of the system of FIG. 1;

FIG. 3 is a block diagram showing the sending of messages from a sender processor to a receiver processor in the system of FIG. 1;

FIG. 4 is a block diagram showing the use of counters associated with buffers of FIG. 3 for monitoring communication events in the system of FIGS. 1 and 3; and

FIG. 5 is thread model showing the creation of a second thread from the user application thread to process notification events and invoke the LAPI handlers of the present invention.

FIG. 1 is a schematic diagram of a distributed computer system 100 useable with the present invention. The distributed computer system 100 may be an IBM RISC System/6000 Scalable POWERparallel Systems (SP) distributed computer system available from International Business Machines Corporation of Armonk, N.Y. The embodiment disclosed in FIG. 1 is an SP computer having a total of 8 frames, with each frame having up to 16 nodes, for a total of 128 nodes. All of the nodes 106 are joined by a local area network (LAN) 102. Each node 106 is a computer itself, and may be a RISC System/6000 workstation, as is well known by those skilled in the art.

All of the nodes in a frame of the SP computer 100 are included in a LAN segment which is joined by the other LAN segments through LAN gates 104. Also connected to the LAN 102, is a control workstation (CWS) 112 which controls operation of the SP computer 100. The control workstation has a direct access storage device (DASD) 114 on which is stored the system data repository files (SDR). The SDR files include such information as a list of nodes that are in the system and their configuration. Each node 106 also includes a DASD device 107 for storing data processed by the SP computer 100.

In one embodiment, the nodes in each frame are also connected to an IBM Scalable POWERparallel switch (SPS) 105. Each of the SPS switches 105 of each frame is connected to neighboring SPS switches 105 of other frames by a bus 110.

As well understood in the art, the CWS node 112 sends system data and control signals to the frames of the SP computer 100 by means of the LAN 102 while messages and data may be sent from one to another of the nodes 106 by means of the high performance switches 105. Adapters (not shown) are placed between each node 106 and the SPS switch 105. One such adapter is shown in U.S. Pat. No. 5,448,558 to Gildea et al. for METHOD AND APPARATUS FOR MANAGING PACKET FIFOS issued Sep. 5, 1995, and owned by the assignee of the present application.

The Low-level Applications Programming Interface (LAPI) of the present invention is a new communications API. As shown in FIG. 2, LAPI is a thin layer above the Hardware Abstraction Layer (HAL) interface and provides reliability through flow control (the HAL interface is disclosed in U.S. patent application Ser. No. 08/920,084 filed Aug. 26, 1997 for "HARDWARE INTERFACE BETWEEN A SWITCH ADAPTER AND A COMMUNICATIONS SUBSYSTEM IN A DATA PROCESSING SYSTEM by C. A. Bender et al. owned by the assignee of the present inventions still pending. LAPI uses a basic "Active Message Style" mechanism which provides a one-sided communications model. The library provides a relatively small set of fixed functions (e.g. LAPI.sub.-- Get, LAPI.sub.-- Put) and a general Active Message function which allows users to supply extensions by means of additions to the notification handlers.

FIG. 3 is a block diagram illustrating the communication low-level application programming interface of the present invention. As shown in FIG. 3, the present embodiment includes a sender or source processor 401 and a receiver or target processor 402. The sender processor 401 is connected to sender or origin user buffers 404A to 404N from which user messages are sent. The sending of user messages from one or the origin user buffers 404A to 404N may be controlled, when present, by a fence 415, to be discussed. A copy of the user message is then packetized in a network buffer 405. An active message passing protocol message handler 406 transfers the user message in packets from the origin network buffer 405 to a receiver or target network buffer 408. LAPI is informed when a message is received in the receiver network buffer 408. LAPI then invokes one of the header handler routines 409A to 409X. The proper header handler routine 409A to 409X to be invoked is indicated in each packet of the message handler by, for instance, a name or address pointing to the location of the header handler routine. The header handler then indicates to LAPI which one of receiver or target user buffers 410A to 410W the user message data is to be stored in, and further indicates which of the completion routines 412A to 412M is to complete the transfer of the user message. LAPI then transfers the user message data from the target network buffer 408 to the indicated target user buffer 410. When the last user message packet is received, LAPI queues the indicated completion routine 412 in one of one or more completion handler queues which, when executed, completes the transfer. The target user buffer may be a null buffer when the message is a command or other message that does not require a target buffer. Similarly, the completion routine may be a null routine when the transfer is completed by the header handler.

The sender processor 401 and receiver processor 402 may each be, for instance, one of the nodes 106 of FIG. 1. It will be understood that messages may be transmitted in either direction between nodes. For example, a message may be sent from the sender 401 to the receiver 402. Also, however, a message may be sent from 402 to 401, in which case 402 becomes the sender and 401 is the receiver.

The completion routines may be tailored to be executed by the receiver processor 402. For instance, even though in the present embodiment, all of the work stations of the nodes 106 are typically the same, a system is envisioned in which the nodes may be different workstations, or even different data processing devices having differing operating systems. In that case, the completion routine is tailored for the particular processor and operating system present at the receiver, and may vary from node to node. Also, since different completion routines, which may be application programs for processing the message data, are present at the receiver, a routine 413 is provided to add to or increase the completion routines 412A to 412M at the receiver, as desired. Similarly, a header handler adder mechanism 414 is provided to add to or increase the header handler routines 409A to 409X at the receiver, as desired. As will be explained, part of the initialization routine for LAPI is to exchange addresses from the receiver to the sender.

In the present embodiment, the header handler indicates the completion routine, if any, which processes the message data. This may be done by including in the header handler, a name of the completion routine desired, an address where the completion routine may be found by the receiver processor, or other means, as is well known to those skilled in the art.

If more than one user message is consecutively transmitted from the sender 401 to the receiver 402, the user messages may arrive at the receiver 402 over the network in any order. A fence 414 may be included in the sender, and the user messages are divided into sets. The set being handled by the message handler 406 is indicated to the fence 414, such as by tokens. If a user message to be sent does not have the proper token, it is not sent until after all of the user messages in the present set have been transmitted and received. In this way, the user messages in the present set are transmitted and placed in the correct user buffer 410, regardless of the order in which the user messages are received.

All user messages are received in the network buffer 408. The receiver user messages may then be placed in one or more user buffers 410, and the user message data is processed by an indicated one of the completion routines at 412, as discussed. The network buffer 408, user buffers 410A to 410W and internal resources, to include completion handler queues, various tables, counters and buffering, may be referred to as receiver resources. If any of the receiver resources are unable to accommodate additional user messages, the sending of additional user messages must be suspended until the receiver resources are again able to handle more user messages. This is done by throttling.

LAPI solves the problem for the case where the rate at which completion routines 412 are started up may be faster than the rate at which they complete. In this situation any buffering (in internal resources) that is provided is bound to get filled. LAPI solves this problem by throttling the sender 401 of these messages. This is done by throttling the sender 401 to not send acknowledgments back for the last network packet of a message that requires a completion routine 412. Thus, the sender 401 cannot complete this message and will resend the unacknowledged packet. When the resent packet arrives, the receiver 402 will check if the required buffer space is now available. If there is available space, the completion routine 412 will be scheduled and the incoming packet acknowledged. If buffering space is still not available, the packet is again ignored (i.e. an acknowledgement is not sent). Thus, the recovery mechanism is used to solve the buffering problem of the completion handler queue.

It will be understood that the acknowledgment (which is an internal message used to control message passing) is sent outside of the critical path of the message passing operation. The user message transmission process, from the start of sending a user message from the sender to the receiver, the receipt of the user message in the network buffer 408, and the placing of the user message in a user buffer 410 by the header handler, has priority over the transmission of internal messages such that user messages do not have to wait for the transmission of internal messages.

Since the processor 402 may also act as a sender to send user messages to the processor 401, if a user message is to be sent to processor 401, and an internal message also has to be sent from processor 402 to 401, some of the internal message may be included with, or "piggybacked", with user messages to reduce the number of messages being sent from node to node.

Since the handling of internal messages causes an interrupt which is disruptive to the operation of the processors of the nodes, the present invention provides a mechanism to increase the number of internal messages that may be handled by a single interrupt. When the sender generates an operation by sending a message, the sender will know how many internal messages will be returned as a result of the operation being generated. Thus, the sender sets a threshold for the message handler 406 which does not raise an interrupt until the expected number of internal messages have been returned. In this way, the number of interrupts are reduced, and a number of internal messages may be handled by a single interrupt.

The present invention uses counters to signal the events associated with the nonblocking communication calls. The specified counter is incremented whenever the event associated with the counter occurs. By checking the value of the counter the user can determine if the event (associated with the counter by the user) in question has occurred.

In the present invention, interfaces that involve nonblocking communication take in as parameters different counters to signal events associated with the communication call in question. The counter structure is defined, and the user is expected to access the counter only using the defined interfaces. However, the user specifies the counters and the relationship of a counter to a message/event. The user may use the same counter across multiple messages. This gives the user the freedom to group different communication calls with the same counter and check their completion as a group.

A library is provided which updates the user specified counters when a particular event (or one of the events) with which the counter was associated has occurred. The user can check on the status of a particular event by polling on the value of the appropriate counter; this polling is done using a LAPI specified interface. Alternatively, the user registers a handler associated with the counter which will be invoked when the counter is incremented or reaches a specified value. If a user wants to wait on a particular event rather than check if it has occurred, LAPI provides an interface to wait on the counter until the counter reaches the specified value. This is a blocking interface. If the user waits for the counter to reach a specific value, then when LAPI returns from the wait, the counter value is decremented by the value (which indicates the number of events) specified in the wait. The same counter may be associated with multiple messages.

By providing the counter solution, a single interface representing the MPI standard and synchronous calls are provided. However, the ability to specify different types of nonblocking behavior is still maintained. The user can simulate either standard or synchronous behavior by specifying appropriate counters and waiting on them at the appropriate places in the code. And the user can mix the use as needed by the application making this more flexible.

FIG. 4 is a block diagram showing an origin 450 and target 451 in a communication system. The origin 450 of FIG. 4 may be the same as the sender 401 of FIG. 3, and the target 451 of FIG. 4 may be the same as the receiver 402 of FIG. 3. A plurality of counters is provided in both the origin 450 and the target 451. These counters may be associated by the user with desired communication events such that the user may monitor the communication events by watching the number in the counters. For instance, one of the counters (shown in FIG. 4 as origin counter 454) may by associated with an origin buffer 453. Another of the counters 455 may be associated with a target buffer 456. The origin buffer 453 may be one of the origin user buffers 404 of FIG. 3, and the target buffer 456 may be one of the target user buffers 412 of FIG. 3. The communication event to be monitored may be sending a message from origin buffer 453 to the target 456, as shown by the arrow 458. Thus, each time a message is sent from origin buffer 453, the origin counter 454 is incremented, and each time a message is received at target 456, the target buffer 456 is incremented. Thus, each time the origin counter 454 is incremented, the user knows, for instance, that the origin buffer 453 is available. Similarly, each time the target counter 455 is incremented, the user knows that the operation on the target buffer 456 by LAPI has been completed.

It will be understood that a number of messages may be grouped together, if desired, and an associated counter may be monitored to determine if the entire group of messages has been transmitted.

A counter 460 may also be defined as a completion counter by the user. When the user message has been completed (as described herein) at the target 451 by the receiver, the target 451 sends a completion message to the origin 450 indicating the message is complete, as shown at 462. As is described, the completion message 462 instructs the completion counter 460 to be incremented. When the completion counter 460 is thus incremented, the origin 450 knows that the message 458 is completed. The completion counter 460 may be used to insure a first message is completed before a second message is transmitted. It will be understood that messages may be delivered in any order. It may be important that a first message be completed before a second message is started. However, the second message may arrive first over the network. In the present embodiment, a user at the origin may send a first message, as shown an 458, and then monitor the completion counter 460. When the completion counter 460 is incremented, the user knows that the first message was completed, and may then issue the second message. This is particularly useful where many messages are being sent, but the order of completion is important for only a few. The order of the desired messages may be assured while allowing those other messages whose order is not important be communicated without delay.

The LAPI is a non-standard application programming interface designed to provide optimal communication performance on the SP Switch. It is based on an "active message style" programming mechanism that provides a one-sided communications model (that is, one process initiates an operation and the completion of that operation does not require any other process to take a complementary action). The LAPI library provides PUT and GET function and a general "active message" function that allows programmers to supply extensions by means of additions to the notification handlers. The LAPI is designed for use by libraries and power programmers for whom performance is more important than code portability.

The LAPI provides the following advantages:

Performance--The LAPI provides basic function for optimal performance. It is designed to especially provide low latency on short messages.

Flexibility--The LAPI's one-sided communications model provides flexibility because the completion of an operation by one process does not require any other process to take a complementary action. Also, the LAPI provides a more primitive interface (than either the MPI or IP) to the SP Switch, giving the programmer the choice of how much additional communications protocol needs to be added.

Extendibility--The LAPI supports programmer-defined handlers that are invoked when a message arrives. Programmers can customize the LAPI to their specific environments.

General characteristics provided by LAPI are:

Reliability--The LAPI provides guaranteed delivery of messages. Errors not directly related to the application are not propagated back to the application.

Flow control

Support for large messages

Non-blocking calls

Interrupt and polling modes

Efficient exploitation of switch function

By default, ordering is not guaranteed

General LAPI Functions

LAPI functions are divided into three parts:

1. A basic "active message" infrastructure that allows programmers to install a set of handlers that are invoked and executed in the address space of a target process on behalf of the process originating the active message. This generic interface allows programmers to customize the LAPI function to their unique environment.

2. A set of defined functions that is complete enough to satisfy the requirements of most programmers. These defined functions make the LAPI more usable and at the same time lend themselves to efficient implementation because their syntax and semantics are known.

3. A set of control functions for the initialization and eventual orderly shutdown of the LAPI layer.

Understanding the LAPI

To help you achieve a fuller understanding of the LAPI, this section presents further details on the "active message" infrastructure and the defined set of functions. In addition, concepts important to understanding the LAPI are explained.

The "Active Message" Infrastructure

"Active message" was selected as the underlying infrastructure for the LAPI. The "active message" infrastructure has the following characteristics:

The "active message" includes the address of a user-specified handler. When the active message arrives at the target process, the specified handler is invoked and executes in the address space of the target process.

The "active message" optionally may also bring with it the user header and data from the originating process.

Operations are unilateral in the sense that the target process does not have to take explicit action for the active message to complete.

Storage buffers for arriving data need to be provided by the invoked handler.

Writing Handlers

The ability for programmers to write their own handlers provides a generalized, yet efficient, mechanism for customizing the interface to one's specific requirements. The user is responsible for protecting shared structures and buffers where necessary by using the locking structures available in the AIX (the IBM version of the UNIX operating system) p-threads library.

The LAPI supports messages that can be larger than the size supported by the underlying SP Switch subsystem. Therefore, the data sent with the "active message" may arrive at the target in multiple packets and, further, these packets can arrive out of order. This situation places some requirements on how the handler is written.

When the "active message" brings with it data from the originating process, the architecture requires that the handler be written as two separate routines:

1. A "header handler" function. This function is specified in the "active message" call. It is called when the message first arrives at the target process and provides the LAPI dispatcher (the part of the LAPI that deals with the arrival of messages and invocation of handlers) with an address where the arriving data must be copied, the address of the optional "completion handler," and a pointer to the parameter that will be passed to the "completion handler".

2. A "completion handler" that is called after the whole message has been received.

An Example of LAPI "Active Message" Function

In this example, a programmer writes a handler for the LAPI "active message" interface. Refer to IBM Parallel System Support Programs for AIX: Command and Technical Reference for more information on the LAPI.sub.-- Amsend subroutine.

1. The desired function (accumulate) is to add vector (S) to another (D) on the target node and put the results in the vector at the target:

where,

S[N] is a vector of length N in the address space of the origin process (origin.sub.-- process)

D[N] is a vector of length N in the address space of the target process (target.sub.-- process)

2. The generic "active message" call is defined as LAPI.sub.-- Amsend (hndl, tgt, hdr.sub.-- hdl, uhdr, uhdr.sub.-- len, udata, udata.sub.-- len, tgt.sub.-- cntr, org.sub.-- cntr, cmpl.sub.-- cntr)

3. Before making the active message call, you must obtain the address of the target counter (target.sub.-- cntr.sub.-- addr) and the address of the header handler to be executed on the target process (accumulate.sub.-- addr). The address of the header handler is obtained by the LAPI.sub.-- Address.sub.-- init function.

4. Initialize the udhr based on the header expected by accumulate. For example, the structure of udhr could be:

5. Make the specific call

6. When this message is received at the target (assuming that the entire origin data is contained within a packet), the accumulate handler you specified is invoked by the dispatcher. The structure of the header handler is:

The structure of the completion handler is:

7. If any state information about the message is required by the completion handler, the information required must be saved in a user buffer by the header handler. The header handler passes the address of this buffer to the dispatcher through the parameter user info. The dispatcher uses this pointer as a parameter (user.sub.-- info) for the completion handler.

8. For this example operations performed at the target process are:

Within the dispatcher:

a. The LAPI header is read.

b. uhdr and uhdr.sub.-- len are extracted from the LAPI header.

c. The header handler is invoked.

d. udata is copied into buf.

e. The completion handler is invoked.

Note: If the message was not contained within a packet, the LAPI layer will save the necessary information and will invoke the completion handler after all the udata has arrived and copied into buf User defined functions:

Header handler:

Completion handler:

The accumulate handler is the header handler and is called by the LAPI layer when the message first arrives at the target process. The header handler saves the information required by complete.sub.-- accumulate (target--addr, length, and buf) in user.sub.-- info and passes back a pointer to the complete.sub.-- accumulate handler as user.sub.-- info. Additionally, the header handler returns address of a buffer buf.

Large active messages are generally transferred as multiple packets. In this case the LAPI layer stores the incoming data in the packets as they arrive into buf. When all the data has been received, it calls the complete.sub.-- accumulate function which uses user.sub.-- info to access the two vectors, adds them and stores them at the desired location. After the return from the complete.sub.-- accumulate routine, the LAPI layer increments tgt.sub.-- ctr. The origin.sub.-- cntr is incremented when it is safe to return the origin buffer back to the user.

The cmpl.sub.-- cntr is incremented after the completion handler has completed execution. The cmpl.sub.-- cntr, therefore, is a reflection, at the origin, of the tgt.sub.-- cntr.

The Defined Set of Functions

Fundamentally, the defined set of functions for the LAPI provides a Remote Memory Copy (RMC) interface. The primary characteristics of the defined set of functions provided by LAPI are:

The basic data transfer operations are memory to memory copy operations that transfer data from one virtual address space to another virtual address space.

The operations are unilateral. That is, one process initiates an operation and the completion of the operation does not require any other process to take some complementary action. (This is unlike a send and receive operation, where a send requires a complementary receive with matching parameters to be posted for completion.)

The operations support both "pull" and "push". The LAPI.sub.-- Get operation copies data from the address space of the target process into the address space of the origin process. The LAPI.sub.-- Put operation copies data into the address space of the target process from the address space of the origin process.

The initiating process specifies the virtual address of both the source and destination of the data (unlike a send and receive process where each side specifies the address in its own address space). To avoid the limitation of requiring that the address maps on the different processes be identical, the LAPI provides the LAPI.sub.-- Address.sub.-- init mechanism by which the different communicating processes can exchange information regarding the address map of shared data objects.

Because data transfer operations are unilateral and no synchronization between the two processes is implied, additional primitives are provided for explicit process synchronization when it is necessary for program correctness.

Functions are provided to detect completion and to enforce ordering.

Important LAPI Concepts

To use the LAPI, it is important to understand the following concepts:

Origin and target

Blocking and non-blocking calls

Completion of communication operation

Message ordering and atomicity

Error handling

Progress

Origin and Target

"Origin" denotes the task (or process or processor) that initiates a LAPI operation (PUT, GET, or "active message".). "Target" denotes the other task whose address space is accessed. Although multiple tasks may run on a single node, it is convenient to think of each task as running on a different node. Therefore the origin task may also be referred to as the "origin node" and the target task as the "target node". The origin and target can be the same for any of the calls, but if the origin and target data areas overlap, the results are undefined.

Blocking and Non-Blocking Calls

A blocking procedure is one that returns only after the operation is complete. There are no restrictions on the reuse of user resources.

A non-blocking procedure is one that may return before the operation is complete and before the user is allowed to reuse all the resources specified in the call. A non-blocking operation is considered to be complete only after a completion testing function, such as LAPI.sub.-- Waitcntr or LAPI.sub.-- Getcntr, indicates that the operation is complete.

Completion of Communication Operation

A communication operation is considered to be complete, with respect to the buffer, when the buffer is reusable.

A PUT is complete with respect to the origin buffer when the data has been copied out of the buffer at the origin and may be overwritten. A GET is complete with respect to the origin buffer when that origin buffer holds the new data that was obtained by GET.

A PUT is complete with respect to the target buffer when the new data is available at the target buffer. A GET is complete with respect to the target buffer when the data has been copied out of the buffer at target and the target task may overwrite that buffer.

Communication Behaviors: Two communication behaviors support two different definitions of "completion":

In standard behavior, a communication operation is defined as complete at the origin task when it is complete with respect to the origin buffer; it is complete at the target task when it is complete with respect to the target buffer.

In synchronous behavior, a communication operation is defined as complete at the origin task when it is complete with respect to both the origin buffer and target buffer. It is complete at the target task when it is complete with respect to the target buffer.

The LAPI defines both standard and synchronous behaviors for PUT operations. The LAPI defines only synchronous behavior for GET operations.

Message Ordering and Atomicity

Two LAPI operations that have the same origin task are considered to be ordered with respect to the origin if one of the operations starts after the other has completed at the origin task. Similarly, two LAPI operations that have the same target task are considered to be ordered with respect to the target if one of the operations starts after the other has completed at the target task. If two operations are not ordered, they are considered concurrent. The LAPI provides no guarantees of ordering for concurrent communication operations. The LAPI does provide mechanisms which an application can use to guarantee order.

As an example, consider the case where a node issues two standard behavior PUT operations to the same target node, where the targets overlap. These two operations may complete in any order, including the possibility of the first PUT overlapping the second, in time. The contents of the overlapping region will be undefined, even after both PUTs complete. Using synchronous behavior for both PUT operations, (waiting for the first to complete before starting the second) will ensure that the overlapping region contains the result of the second after both PUTs have completed.

Error Handling

If an error occurs during a communications operation, the error may be signaled at the origin of operation, or the target or both. Some errors may be caught before the communication operation begins, and these will be signaled at the origin. However, some errors will not occur until the communication is in progress (a segmentation violation at the target, for example); these may be signaled at either or both ends of the communication.

Progress

All LAPI operations are unilateral by default and can complete successfully or fail, independent of the actions of other tasks. Specifically, a LAPI operation to a particular target should complete even if the target is stuck in an infinite loop (that is, when the target process is in interrupt mode).

Using the LAPI

In general, LAPI functions:

Are non-blocking calls

Provide both polling and interrupt mode

Signal completion by incrementing counters at each end

Provide both C and Fortran bindings

Complementary functions provide for checking completion of operations and for enforcing relative ordering if required. Additionally, functions allow processes to exchange addresses that will be used in LAPI operations.

Specific LAPI Functions

The LAPI provides the following specific functions.

Active Message Active message function (LAPI.sub.-- Amsend) is a non-blocking call that causes the specified active message handler to be invoked and executed in the address space of the target process. Completion of the operation is signaled if counters are specified. Both standard and synchronous behaviors are supported. The LAPI.sub.-- Amsend function provides two counters (org.sub.-- cntr and cmpl.sub.-- cntr) which can be used to provide the two behaviors. The org.sub.-- cntr is incremented when the origin buffer can be reused (standard). The cmpl.sub.-- cntr is incremented after the completion handler has completed execution (synchronous).

Data Transfer

Data transfer functions are non-blocking calls that cause data to be copied from a specified region in the origin address space to the specified region in the target address space (in the case of a LAPI.sub.-- Put operation) or from a specified region in the target address space to a specified region in the origin address space (in the case of a LAPI.sub.-- Get operation). Completion of the operation is signaled if counters are specified. Both standard and synchronous operations are supported for PUT. Only synchronous operation is possible in the case of GET. Standard PUT is provided by incrementing the org.sub.-- cntr when the origin buffer can be reused.

Synchronous PUT is provided by incrementing the cmpl.sub.-- cntr after the data has been written into the target buffer. The LAPI.sub.-- Getcntr (or LAPI.sub.-- Waitcntr) function should be used in conjunction with the org.sub.-- cntr and cmpl.sub.-- cntr counters to guarantee the respective standard and synchronous behavior of the LAPI PUT operations.

Synchronizing

The LAPI.sub.-- Rmw function is used to synchronize two independent operations such as two processes sharing a common data structure. The operation is performed at the target process and is atomic. The operation takes a variable from the origin and performs one of four selected operations on a variable from the target and replaces the target variable with the results of the operation. The original value of the target variable is returned to the origin. LAPI.sub.-- Rmw provides four different read/modify/write (rmw) operations:

SWAP

COMPARE.sub.-- AND.sub.-- SWAP

FETCH.sub.-- AND.sub.-- ADD

FETCH.sub.-- AND.sub.-- OR

Completion is signaled at the origin if the counter is specified.

Completion Checking

The following counter functions provide the means for a process to manage the completion state of the LAPI operations.

LAPI.sub.-- Waitcntr--Wait on a counter to reach a specified value and return when the counter is equal to or greater than that value (blocking)

LAPI.sub.-- Getcntr--Get the current value of a specified counter (non-blocking)

LAPI.sub.-- Setcntr--Set the counter to a specified value

These functions also provide an efficient means to order the flow of LAPI operations or the use of certain user managed resources (for example, buffers). For example, a series of PUTs to a single target and buffer requires that the contents of the buffer at the target remains in step with the order of execution of the PUTs at the origin. Using the cmpl.sub.-- cntr counter in the LAPI.sub.-- Put function in conjunction with the LAPI.sub.-- Waitcntr function provides the necessary ordering.

Ordering

LAPI.sub.-- Fence and LAPI.sub.-- Gfence operations provide a means to enforce the order of execution of LAPI functions. LAPI functions initiated prior to these fencing operations are guaranteed to complete before LAPI functions initiated after the fencing functions. LAPI.sub.-- Fence is a local operation which is used to guarantee that all LAPI operations initiated by the local process and the same process thread are complete. LAPI.sub.-- Gfence is a collective operation involving all processes in the parallel program. LAPI.sub.-- Gfence provides a barrier operation for the parallel program. Both LAPI.sub.-- Fence and LAPI.sub.-- Gfence operations are a data fence that guarantee that the data movement is complete. These are not an operation fence which would need to include "active message" completion handlers completing at the target.

Progress

The LAPI.sub.-- Probe function is used in polling mode to transfer control to the communication subsystem in order to make progress on arriving messages.

Address Manipulation

The LAPI.sub.-- Address.sub.-- init collective operation allows processes to exchange operand addresses of interest. Such function is required if the processes do not have identical address maps. The LAPI.sub.-- Address is used by Fortran programs when an address needs to be stored in an array. In Fortran there is no concept of "address" (as there is in the C language) and this function gives that ability to Fortran.

LAPI Setup

LAPI.sub.-- Init and LAPI.sub.-- Term operations are used to initialize and terminate the communication structures required to effect LAPI communications for a specific instance of LAPI. LAPI.sub.-- Init returns a unique handle which is used to associate a communication channel with the LAPI instance. This handle is subsequently passed as a parameter to each of the other LAPI functions. The LAPI.sub.-- Term function is used to terminate a specific instance of LAPI.

Error Handling and Messages

The LAPI.sub.-- Init function provides a means for the user of LAPI to register an error handler. This error handler is specified as part of the lapi.sub.-- info structure parameter which is passed to the LAPI.sub.-- Init function. The LAPI.sub.-- Msg.sub.-- String function provides the means to translate a LAPI call return code value (integer) into a message string.

LAPI Environment

The LAPI.sub.-- Qenv function is used to query the state of the LAPI communications subsystem. The LAPI.sub.-- Senv function allows the programmer to specify the value of some of the LAPI communications subsystem's environment variables. An important value that can be specified is the interrupt state. The interrupt state is set by specifying INTERRUPT.sub.-- SET as on (for interrupt mode) or off (for polling mode). The default setting for INTERRUPT.sub.-- SET is on.

The LAPI Execution Model

The goal of LAPI is to provide a threads-safe environment and support an execution model that allows for maximum execution concurrency within the LAPI library.

Using the setup function (LAPI.sub.-- Init), a user process establishes a LAPI context. Within a LAPI context, the LAPI library is threads-safe, and multiple threads may make LAPI calls within the same context. The different calls may execute concurrently with each other and with the user threads. However, in reality execution concurrency among these calls is limited by the locking required with LAPI to maintain integrity of its internal data structures and the need to share a single underlying communication channel.

As with any multi-threaded application, coherence of user data is the responsibility of the user. Specifically, if two or more LAPI calls from different threads can execute concurrently and if they specify overlapping user buffer areas, then the result is undefined. It is the responsibility of the user to coordinate the required synchronization between threads that operate on overlapping buffers.

The user application thread, as well as the completion handlers, cannot hold mutual exclusion resources before making LAPI calls; if they do, it is possible to run into deadlock situations.

Because user-defined handlers may be called concurrently from multiple threads, it is the user's responsibility to make them threads-safe.

From the user application thread a second thread is created by LAPI.sub.-- Init to process notification events and invoke the LAPI handlers as illustrated in FIG. 5.

The application thread, notification thread and the completion handler thread are shown in FIG. 5.

Threads 0 and 1 (the application thread and the notification thread) attempt to invoke the LAPI dispatcher whenever possible; in this way, progress on incoming and outgoing messages can be made while minimizing additional overhead. Most LAPI calls (though not all) made by the application thread also result in the LAPI dispatcher being automatically run. The notification thread waits in the Kernel for the occurrence of a notification event. When an event occurs, the Kernel wakes up the waiting thread. As shown in FIG. 5, after the notification thread returns from waiting in the Kernel, it invokes the LAPI dispatcher.

The LAPI Dispatcher is the central control point that orchestrates the invocation of functions and threads necessary to process outstanding incoming and outgoing LAPI messages.

The LAPI Dispatcher can run from the application's user's thread, from the notification thread or from the completion handler thread. Locking is used to ensure that only one instance of the dispatcher runs at a time to maintain integrity. On incoming messages, the LAPI dispatcher manages the reassembly of data from different packets (which may arrive out-of-order) into the specified buffer, and then invokes the completion handler if necessary.

Thread 2 is created by LAPI.sub.-- Init to execute completion handlers associated with "active messages". Completion handlers are written by users and may make LAPI function calls which in turn will invoke the LAPI Dispatcher. The completion handler thread processes work from the completion handler queue. When the queue is empty the thread waits using a pthread.sub.-- cond.sub.-- wait(). If an active message (LAPI.sub.-- Amsend) includes a completion handler, the dispatcher queues a request on the completion queue after the whole message has arrived and has been reassembled in the specified buffer; the dispatcher then sends a pthread.sub.-- cond.sub.-- signal to the completion handler thread. If this thread was in a wait state it will begin processing the completion handler queue, otherwise, if it was not waiting, the thread signal is ignored.

LAPI handlers are not guaranteed to execute one at a time. It is important to note, however, that LAPI calls may execute concurrently with the origin or target or both. The same restrictions stated previously about not holding on to mutual exclusion resources when making LAPI calls still applies.

This discussion of a threads-safe environment and maximum execution concurrency within the LAPI library applies to both the polling and interrupt modes. In polling mode any calls to the communication library attempt to make progress on the context specified in the call. Further, the function LAPI.sub.-- Probe is provided to allow applications to explicitly check for and handle incoming messages.

The execution model of the handlers consists of the following events:

Event Action

Message Arrival

Copies the message from the network into the appropriate data access memory space.

Interrupt/Poll

Causes an interrupt if required, based on the mode.

Dispatcher Start

Invokes the dispatcher.

New Message Packet

Checks the LAPI header and determines (by checking the receive state message reassembly table) if the packet is part of a pending message or if it is a new message. Calls the header-handler function.

Return from Header-Handler

If the message is contained in more than one packet, the LAPI Dispatcher will log that there is a pending message, save the completion handler address, and save the user's buffer address to be used during the message reassembly of pending message packets.

Pending Message Packet

Copies the message to the appropriate portion of the user buffer specified through the header-handler. If the packet completes the message, the dispatcher queues the completion handler; otherwise the dispatcher returns to check for message arrivals.

Completion Handler

When the completion handler is executed, (after the return from the completion handler) updates the appropriate target counter before continuing.

Allocating Buffers

1. The user allocates as many buffers per origin as wanted.

Rule: At origin, if the target buffer allocation is exhausted, wait for the counter of previous requests.

Example: Use>=P buffers (one per origin). At origin with one request pending. Wait on counter before issuing next request.

2. If the header handler blocks, no further progress is made, including messages pending (that is, the communications adapter is stalled).

When the order of execution between any two LAPI functions within one task of a parallel program needs to be guaranteed, using LAPI.sub.-- Waitcntr between the two LAPI functions will usually be more efficient than LAPI.sub.-- Fence. LAPI.sub.-- Fence requires that all LAPI operations initiated on the current thread before the LAPI.sub.-- Fence be completed before any LAPI operation after the fence is allowed to start. LAPI.sub.-- Waitcntr can be used to indicate the completion of a single LAPI function which had been initiated on the current thread before the LAPI.sub.-- Waitcntr.

The scope of LAPI.sub.-- Fence is per thread. For example, a LAPI.sub.-- Fence which is issued from the completion handler thread will only guarantee that no LAPI operations initiated after the fence (on the completion handler thread) will start until all LAPI operations initiated before the fence have completed. In this case there are no guarantees about the order of LAPI operations initiated from the main application thread.

LAPI.sub.-- Waitcntr can be used to indicate the completion of a single LAPI function which may have been initiated from an alternate thread (completion handler) within the same task. Therefore the possibility exists to use LAPI.sub.-- Waitcntr to wait for the completion of another LAPI function which is initiated after the call to LAPI.sub.-- Waitcntr.

LAPI.sub.-- Waitcntr can be used to guarantee order of execution of LAPI.sub.-- Amsend operations which are initiated from a single origin task. When LAPI.sub.-- Amsend operations use the cmpl.sub.-- cntr counter, this counter is incremented after the completion counter (or header handler if a completion handler is not specified) has executed at the target task. LAPI.sub.-- Fence and LAPI.sub.-- Gfence do not provide an indication that LAPI.sub.-- Amsend operations have completed execution at the target.

LAPI.sub.-- Waitcntr is a blocking call. If a user prefers to avoid this blocking operation a program loop comprised of the sequence

LAPI.sub.-- Getcntr

a check of the value returned from Get

LAPI.sub.-- Probe

will provide an equivalent logical operation and provide the user with added flexibility.

LAPI.sub.-- Init must be called before any thread (the main thread or the completion handler thread) can make a LAPI call. In addition to this, LAPI.sub.-- Address.sub.-- init or LAPI.sub.-- Gfence should be the second LAPI call. These two functions provide a barrier which guarantees that all other LAPI tasks have initialized their LAPI subsystems and are ready to receive requests from remote tasks. Failure to provide these barrier functions may result in dropped switch packets, low performance at start-up and unnecessary switch congestion. The instance of the LAPI subsystem should be quiesced before LAPI.sub.-- Term is called to terminate the LAPI instance. This can be done by calling LAPI.sub.-- Gfence before LAPI.sub.-- Term.

LAPI Programming Example

Active Message (C)

This C program is an example of the use of the LAPI

active message call:

Following is a list of subroutines and their definitions usable with LAPI of the present invention. These subroutine definitions will be understood by those skilled in the art for use in the present implementation of LAPI.

Initialization and Termination

Initialize LAPI

LAPI.sub.-- Init, LAPI.sub.-- INIT Function

Purpose

Initialize the LAPI subsystem.

C Syntax

Parameters

Description

This function instantiates a new context of LAPI subsystem and initializes it. A handle to the newly created LAPI context is returned in hndl. All subsequent LAPI calls can use hndl to specify the context of the LAPI operation.

lapi.sub.-- dev.sub.-- t is defined as follows:

NOTE: Only TB3.sub.-- DEV lapi.sub.-- dev.sub.-- t type is suppoted in this release. User can register an error handler through lapi.sub.-- info structure. The user needs to create a function with the following parameters:

Error code(*error.sub.-- code) of LAPI.sub.-- ERR.sub.-- TIMEOUT is a recoverable error if the user chooses to ignore it in the user's error handler. All other error codes are currently terminal and the user should do clean up processing of user's process and terminate the process (i.e. exit()).

Except for LAPI.sub.-- Address() and LAPI.sub.-- Msg.sub.-- string(), it is an error to make any LAPI calls before calling LAPI.sub.-- Init().

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

EBUSY--System error--Previous job still running.

EINVAL--System error--Invalid argument.

EPERM--System error--caller not authorized to perform the action.

ETIMEDOUT--System error--Switch network is not up.

ENODEV--System error--Adapter type and library do not match.

ENOSPC--System error--Can not attach to bus memory

out of memory or segment register.

CSS.sub.-- KE.sub.-- INTERNAL.sub.-- ERROR--System error--Kernel extension internal memory management failed.

CSS.sub.-- KE.sub.-- UCODE.sub.-- ERROR--System error--Adapter micro code is not responding.

LAPI.sub.-- ERR.sub.-- UNKNOWN.sub.-- DEVICE--Not supported device.

LAPI.sub.-- ERR.sub.-- NOMORE.sub.-- PORTS--No more communication ports are available.

LAPI.sub.-- ERR.sub.-- INIT.sub.-- FAILED--Initialization failed.

LAPI.sub.-- ERR.sub.-- OPEN.sub.-- FAILED--Opening of communication device failed.

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Term.

Terminate the LAPI context

LAPI.sub.-- Term, LAPI.sub.-- TERM Function

Purpose

Terminate and cleanup the LAPI subsystem.

C Syntax

Parameters

Description

This function terminates the LAPI context specified by hndl. Any LAPI notification threads associated with this context will be terminated. It is an error to make any LAPI calls using hndl after LAPI.sub.-- Term() has been called, except for LAPI.sub.-- Msg.sub.-- string() and LAPI.sub.-- Address().

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

EINVAL--System error--Invalid argument.

EPERM--System error--caller not authorized to perform the action.

LAPI.sub.-- ERR.sub.-- CLOSE.sub.-- FAILED--Close of communication device failed.

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Init.

Return Message string, Query and Set LAPI

Environment Functions

Get LAPI message string.

LAPI.sub.-- Msg.sub.-- string, LAPI.sub.-- MSG.sub.-- STRING Function

Purpose

Get LAPI and System Message String.

C Syntax

Parameters

Description

This function returns the message string representation of the return value for a specific LAPI call.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Init, LAPI.sub.-- Term.

Query LAPI Environment.

LAPI.sub.-- Qenv, LAPI.sub.-- QENV Function

Purpose

Query the LAPI interface for parallel job information.

C Syntax

Parameters

Description

This function queries the LAPI interface for information about a specific LAPI instance.

lapi.sub.-- query.sub.-- t defines the types of LAPI queries available.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Amsend, LAPI.sub.-- Get, LAPI.sub.-- Put, LAPI.sub.-- Senv.

Set LAPI Environment.

LAPI.sub.-- Senv, LAPI.sub.-- SENV Function

Purpose

Set the LAPI environment for given context.

C Syntax

Parameters

Description

This function sets the LAPI environment for specific LAPI instance. lapi.sub.-- query.sub.-- t defines the types of LAPI set environment variables.

NOTE: If ERROR.sub.-- CHK is set to 0 then for all LAPI calls parameter error checking are ignored (i.e. LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER is not returned).

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Qenv.

Basic Data Transfer Calls

Direct Put

LAPI.sub.-- Put, LAPI.sub.-- PUT Function

Purpose

Put data into target address on a target process.

C Syntax

Parameters

Description

This function transfers len number of bytes from virtual address org.sub.-- addr on the origin to the target process tgt at the address tgt.sub.-- addr over the port identified by hndl. After the data has been copied out of the memory at org.sub.-- addr the org.sub.-- cntr is incremented. After the data has arrived at the tgt, the tgt.sub.-- cntr is incremented. If either counter address is NULL, the data transfer happens, but the corresponding counter increment do not take place. This is a non blocking call, in that, the calling program may not assume that origin buffer may be changed, nor that contents of the memory pointed to by tgt.sub.-- addr on tgt is ready for use. However, after the origin waits for the org.sub.-- cntr update to complete, the origin can modify the origin buffer org.sub.-- addr. Similarly, the target can modify the data in the target buffer tgt.sub.-- addr after it has waited for tgt.sub.-- cntr update to complete on the target. This call can be made synchronous if the origin waits for the cmpl.sub.-- cntr update to complete.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Fence, LAPI.sub.-- Get, LAPI.sub.-- Getcntr, LAPI.sub.-- Gfence,

LAPI.sub.-- Qenv, LAPI.sub.-- Waitcntr.

Direct Get

LAPI.sub.-- Get, LAPI.sub.-- GET Function

Purpose

Copy data from a remote process to the local address on local process.

C Syntax

Parameters

Description

Use this subroutine to transfers len number of bytes from tgt.sub.-- addr address at the target process to the org.sub.-- addr virtual address at the origin process over the port identified by hndl. After the data is copied out of the memory at tgt.sub.-- addr, the tgt.sub.-- cntr is incremented. After the data arrives at the origin, the org.sub.-- cntr is incremented. If either counter address is NULL, the data transfer occurs, but the corresponding counter increment does not take place. This is a non blocking call in that the calling program (all processes associated with the program including the local and remote processes involved) can not assume that the target buffer can be changed, nor that contents of the memory pointed to by org.sub.-- addr on the origin is ready for use. However, after the origin waits for the org.sub.-- cntr update to complete, the origin can use the org.sub.-- addr data. Similarly, the target can reuse the target buffer tgt.sub.-- addr only after it has waited for the tgt.sub.-- cntr update to complete at the target.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

None.

Related Information

LAPI.sub.-- Fence, LAPI.sub.-- Getcntr, LAPI.sub.-- Gfence, LAPI.sub.-- Put,

LAPI.sub.-- Qenv, LAPI.sub.-- Waitcntr.

Get, Wait, Set and Probe Operations

Get Counter Operation

LAPI.sub.-- Getcntr, LAPI.sub.-- GETCNTR Function

Purpose

Gets the integer value of counter.

C Syntax

Parameters

Description

Gets the integer value of cntr. This can be used to see how much progress is being made in LAPI context hndl.

In conjunction, LAPI.sub.-- Probe() can be used to make progress in LAPI context hndl if LAPI.sub.-- Getcntr() is called inside a loop.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Amsend, LAPI.sub.-- Get, LAPI.sub.-- Probe, LAPI.sub.-- Put,

LAPI.sub.-- Setcntr, LAPI.sub.-- Waitcntr.

Wait Counter Operation

LAPI.sub.-- Waitcntr, LAPI.sub.-- WAITCNTR Function

Purpose

Waits till a specified counter reaches the value specified.

C Syntax

Parameters

Description

This function waits till the cntr reaches or exceeds the specified val. Once the cntr reaches the val, the cntr is decremented by that value. (we say decremented rather than set to zero since the cntr could has had a value greater than the specified val, when the call was made). This call may or may not check for message arrivals over the LAPI context hndl; LAPI.sub.-- Probe will always check for message arrivals.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Amsend, LAPI.sub.-- Get, LAPI.sub.-- Getcntr, LAPI.sub.-- Probe,

LAPI.sub.-- Put, LAPI.sub.-- Rmw, LAPI.sub.-- Setcntr.

Set Counter Operation

LAPI.sub.-- Setcntr, LAPI.sub.-- SETCNTR Function

Purpose

Sets a counter to a specified value.

C Syntax

Parameters

Description

This function sets the cntr to the appropriate value.

The LAPI context associated with hndl may or may not be polled for incoming messages.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Getcntr, LAPI.sub.-- Probe, LAPI.sub.-- Waitcntr.

LAPI Probe

LAPI.sub.-- Probe, LAPI.sub.-- PROBE Function

Purpose

To transfer control to the communications subsystem to check for arriving messages and to make progress in polling mode.

C Syntax

Parameters

Description

This function transfers control to the communications subsystem in order to make progress on messages associated with the context hndl. Note that no guarantees are made about receipt of messages on the return from this function.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Getcntr, LAPI.sub.-- Waitcntr, LAPI.sub.-- Setcntr.

Read-Modify-Write Operations

LAPI.sub.-- Rmw, LAPI.sub.-- RMW Function

Purpose

This function provides the synchronization primitives.

C Syntax

Parameters

Description

The LAPI.sub.-- Rmw function is used to synchronize two independent operations such as two processes sharing a common data structure. The operation is performed at the tgt target process and is atomic. The operation takes a in.sub.-- val from the origin and performs one of four selected op operations on a tgt.sub.-- var variable at the tgt target and replaces the tgt.sub.-- var target variable with the results of the op operation. The prev.sub.-- tgt.sub.-- val original value of the tgt.sub.-- var target variable is returned to the origin.

The valid operations for op are:

FETCH.sub.-- AND.sub.-- ADD

FETCH.sub.-- AND.sub.-- OR

SWAP

COMPARE.sub.-- AND.sub.-- SWAP.

The operations are performed over the context referred to by hndl. The outcome of the execution of these calls is as if the following code was executed atomically.

where f(a; b)=a+b for FETCH.sub.-- AND.sub.-- ADD,

f(a; b)=a.vertline.b for FETCH.sub.-- AND.sub.-- OR (bitwise or),

and f(a; b)=b for SWAP.

For COMPARE.sub.-- AND.sub.-- SWAP in.sub.-- val is treated as a pointer to an array of two integers, and the op is the following atomic operation.

All the calls are nonblocking. Testing for completion is by using LAPI.sub.-- Getcntr and LAPI.sub.-- Waitcntr functions described above. There is no tgt.sub.-- cntr on RMW calls, and they do not provide any indication of completion on the tgt process.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Getcntr, LAPI.sub.-- Probe, LAPI.sub.-- Qenv,

LAPI.sub.-- Setcntr, LAPI.sub.-- Waitcntr.

Fence Calls

Local fence

LAPI.sub.-- Fence, LAPI.sub.-- FENCE Function

Purpose

This function is used to enforce order on LAPI calls.

C Syntax

Parameters

Description

If a process calls LAPI.sub.-- Fence(), all the LAPI operations that were initiated by that process before the fence using the LAPI context hndl are guaranteed to complete at the target processes before any of its communication operations using hndl initiated after the fence complete at the target processes. This is a data fence which means that the data movement is complete. This is not an operation fence which would need to include Active Message completion handlers completing on the target.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Amsend, LAPI.sub.-- Get, LAPI.sub.-- Gfence, LAPI.sub.-- Put, LAPI.sub.-- Rmw.

Global Fence

LAPI.sub.-- Gfence, LAPI.sub.-- GFENCE Function

Purpose

This function is used to enforce order on LAPI calls on all processes.

C Syntax

Parameters

Description

This is a collective call. On completion of this call, it is assumed that all LAPI communication associated with hndl from all processes has quiesced. Note that although hndl is a local variable, it has a set of nodes that were associated with it at LAPI.sub.-- Init all of which have to participate in this operation for it to complete. This is a data fence which means that the data movement is complete. This is not an operation fence which would need to include Active Message completion handlers completing on the target.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Fence.

Exchanging Virtual Addresses

All LAPI data transfer calls are required to provide virtual addresses on the target processes. How does a process know the virtual addresses of objects on other processes?

In the case of SPMD programs it may be possible to assume that all processes have the same virtual address map. In such cases, one may assume that the local virtual address of a static object (global variables in C or fortran common blocks) is the same as the virtual address of the same object on a different process. This assumption is Operating System dependent, and is true under the current versions of AIX. However, AIX has the ability to relocate static data objects when an executable file is loaded and, this assumption may not be valid on future versions of AIX.

Get and Put can be used to move addresses, but in order to do so, there needs to be a mechanism provided for boot-strapping. If all processes could agree on a specific virtual address where they place an array of pointers to objects of interest, then a process could simply get the appropriate address from this table by indexing with respect to the address of the array of pointers. This approach may not work in non-SPMD programs, where it may not be possible to agree on such a specific address.

To solve this problem, the following function in LAPI is defined, which is a collective communication operation, that basically collects one address from each process and distributes all the addresses to all the processes.

LAPI.sub.-- Address.sub.-- init, LAPI.sub.-- ADDRESS.sub.-- INIT Function

Purpose

To exchange virtual addresses for non SPMD programs and for dynamically allocated data.

C Syntax

Parameters

Description

This function is used to exchange virtual and dynamically allocated addresses. add.sub.-- tab is an array of pointers of size LAPI.sub.-- Qenv(,NUM.sub.-- TASKS,). This function is a collective call over the LAPI context hndl which fills the table add.sub.-- tab with the entries supplied by each task. Upon completion of this call, add.sub.-- tab[i] will contain the entry provided by task i.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

None.

Generic Active Message Functions

The LAPI library provides a generic Active Message function as a means to allow users to extend the base set of LAPI functions. The Active Message function simply provides a means to call a user written handler.

LAPI Active Message

LAPI.sub.-- Amsend, LAPI.sub.-- AMSEND Function

Purpose

To invoke a user provided Active Message (AM) handler to run on a remote (target) process.

C Syntax

Parameters

Description

This function transfers hdr.sub.-- hdl function pointer along with the contents of uhdr and udata from the origin to the target process tgt. When the message arrives at the target process, the header-handler hdr.sub.-- hdl is invoked at the tgt target with the pointer to uhdr as one of the parameters. The user-supplied header-handler is expected to return a buffer pointer (user.sub.-- info) as the return value, in which udata is to be copied. The header-handler is also expected to save any information that will be required later by the completion-handler. The header-handler returns (through reference parameters) the completion-handler and a pointer to the saved information (user.sub.-- info).

Note: that the header-handler should be non-blocking because no progress on the messages associated with hndl can be made until control is returned to the communications library from the header-handler. After the header-handler returns, udata (if any) is copied into the user-specified buffer (user.sub.-- info). When all of the udata has been copied into the user buffer, the completion-handler specified by the user through the header-handler is enqueued. After the parameters (including contents of uhdr and udata) are copied out of the memory at the origin, the org.sub.-- cntr is incremented. After the completion-handler has completed execution at the tgt, the tgt.sub.-- cntr is incremented. If the completion handler specified is a NULL, then tgt.sub.-- cntr is incremented after all of udata has been copied into the user specified buffers. If the user specified buffer is NULL and the completion handler is also NULL, then tgt.sub.-- cntr will be incremented in some implementation specific manner. Either counter addresses may be NULL. This is a non blocking call. The calling process cannot change the uhdr origin header and udata data until completion at the origin is signaled by the org.sub.-- cntr being incremented. Similarly, it can be assumed that the specified AM handler has run at tgt only after the target counter tgt.sub.-- cntr has been incremented. The cmpl.sub.-- cntr and tgt.sub.-- cntr counters will be incremented after the AM handler has completed execution at the target. When the AM handler has both a hdr.sub.-- hdl header handler and comp.sub.-- h completion handler, the cmpl.sub.-- cntr and tgt.sub.-- cntr counters will be incremented after the completion handler has completed execution. If the AM handler has only a hdr.sub.-- hdl header handler, the cmpl.sub.-- cntr and tgt.sub.-- cntr counters will be incremented after the header handler has completed execution. This call can be made synchronous if the origin waits for the cmpl.sub.-- cntr update to complete. The length (uhdr.sub.-- len) of the user specified header is constrained by an implementation specified maximum value (LAPI.sub.-- Qenv(,MAX.sub.-- UHDR.sub.-- SZ,)). In the current implementation, the amount of udata sent per packet is LAPI.sub.-- Qenv(,MAX.sub.-- UHDR.sub.-- SZ,)--uhdr.sub.-- len. To get the best bandwidth uhdr.sub.-- len should be as small as possible.

Return Values

LAPI.sub.-- SUCCESS--on successful completion.

The following can be returned on error:

LAPI.sub.-- ERR.sub.-- BAD.sub.-- PARAMETER--Invalid parameter passed in.

Related Information

LAPI.sub.-- Fence, LAPI.sub.-- Getcntr, LAPI.sub.-- Qenv, LAPI.sub.-- Waitcntr.

While we have illustrated and described the preferred embodiment of our invention, it is to be understood that we do not limit ourselves to the precise construction herein disclosed, and the right is reserved to all changes and modifications coming within the scope of the invention as defined in the appended claims.