Collective communication system and methods

Abstract

A method in which a plurality of process are configured to hold a block of data destined for other processes, with data repacking circuitry including receiving circuitry configured to receive at least one block of data from a source process of the plurality of processes, the repacking circuitry configured to repack received data in accordance with at least one destination process of the plurality of processes, and sending circuitry configured to send the repacked data to the at least one destination process of the plurality of processes, receiving a set of data for all-to-all data exchange, the set of data being configured as a matrix, the matrix being distributed among the plurality of processes, and transposing the data by each of the plurality of processes sending matrix data from the process to the repacking circuitry, and the repacking circuitry receiving, repacking, and sending the resulting matrix data to destination processes.

Description

FIELD OF THE INVENTION

The present invention, in exemplary embodiments thereof, relates to collective communication systems and methods, and particularly but not exclusively to message passing operations, and also particularly but not exclusively to all-to-all operations.

SUMMARY OF THE INVENTION

The present invention, in certain embodiments thereof, seeks to provide improved systems and methods for collective communication, and in particular, but not only, for message passing operations, including all-to-all operations.

There is thus provided in accordance with an exemplary embodiment of the present invention a method including providing a plurality of processes, each of the plurality of process being configured to hold a block of data destined for others of the plurality of processes, providing at least one instance of data repacking circuitry including receiving circuitry configured to receive at least one block of data from at least one source process of the plurality of processes, repacking circuitry configured to repack received data in accordance with at least one destination process of the plurality of processes, and sending circuitry configured to send the repacked data to the at least one destination process of the plurality of processes, receiving a set of data for all-to-all data exchange, the set of data being configured as a matrix, the matrix being distributed among the plurality of processes, and transposing the data by each of the plurality of processes sending matrix data from the process to the repacking circuitry, and the repacking circuitry receiving, repacking, and sending the resulting matrix data to destination processes.

Further in accordance with an exemplary embodiment of the present invention the method also includes providing a control tree configured to control the plurality of processes and the repacking circuitry.

Still further in accordance with an exemplary embodiment of the present invention the control tree is further configured to receive registration messages from each of the plurality of processes, mark a given subgroup of the plurality of processes as ready for operation when registration messages have been received from all members of the given subgroup, when a given subgroup which is a source subgroup and a corresponding subgroup which is a destination subgroup are ready for operation, pair the given source subgroup and the given destination subgroup and assign the given source subgroup and the given destination subgroup to an instance of repacking circuitry, and notify each the source subgroup and each the destination subgroup when operations relating to each the source subgroup and each the destination subgroup have completed.

Additionally in accordance with an exemplary embodiment of the present invention the control tree is configured, in addition to pairing the given source subgroup and the given destination subgroup, to assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry.

Moreover in accordance with an exemplary embodiment of the present invention the method also includes assigning circuitry other than the control tree, the assigning circuitry being configured to assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry.

Further in accordance with an exemplary embodiment of the present invention the control tree includes a reduction tree.

There is also provided in accordance with another exemplary embodiment of the present invention apparatus including receiving circuitry configured to receive at least one block of data from at least one source process of a plurality of processes, each of the plurality of process being configured to hold a block of data destined for others of the plurality of processes, at least one instance of data repacking circuitry configured to repack received data in accordance with at least one destination process of the plurality of processes, and sending circuitry configured to send the repacked data to the at least one destination process of the plurality of processes, the apparatus being configured to receive a set of data for all-to-all data exchange, the set of data being configured as a matrix, the matrix being distributed among the plurality of processes, and the apparatus being further configured to transpose the data by receiving, from each of the plurality of processes, matrix data from the process at the repacking circuitry, and the data repacking circuitry receiving, repacking, and sending the resulting matrix data to destination processes.

Further in accordance with an exemplary embodiment of the present invention the apparatus also includes a control tree configured to control the plurality of processes and the repacking circuitry.

Still further in accordance with an exemplary embodiment of the present invention the control tree is further configured to receive registration messages from each of the plurality of processes, mark a given subgroup of the plurality of processes as ready for operation when registration messages have been received from all members of the given subgroup, when a given subgroup which is a source subgroup and a corresponding subgroup which is a destination subgroup are ready for operation, pair the given source subgroup and the given destination subgroup and assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry, and notify each the source subgroup and each the destination subgroup when operations relating to each source subgroup and each destination subgroup have completed.

Additionally in accordance with an exemplary embodiment of the present invention the control tree is configured, in addition to pairing the given source subgroup and the given destination subgroup, to assign the given source subgroup and the given destination subgroup to a given instance of data repacking circuitry.

Moreover in accordance with an exemplary embodiment of the present invention the apparatus also includes assigning circuitry other than the control tree, the assigning circuitry being configured to assign the given source subgroup and the given destination subgroup to a given instance of data repacking circuitry.

Further in accordance with an exemplary embodiment of the present invention the control tree includes a reduction tree.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be understood and appreciated more fully from the following detailed description, taken in conjunction with the drawings in which:

FIG. 1A is a simplified pictorial illustration of an exemplary computer system, constructed and operative in accordance with an exemplary embodiment of the present invention;

FIG. 1B is a simplified pictorial illustration of an exemplary data block layout;

FIG. 2 is a simplified pictorial illustration of another exemplary data block layout;

FIG. 3 is a simplified pictorial illustration depicting all-to-all-v initial and final stages;

FIG. 4 is a simplified pictorial illustration depicting direct pairwise exchange;

FIG. 5 is a simplified pictorial illustration depicting an aggregation algorithm;

FIG. 6 is a simplified pictorial illustration depicting initial block distribution for an all-to-all operation, in accordance with an exemplary embodiment of the present invention;

FIG. 7 is a simplified pictorial illustration depicting final block distribution for an all-to-all operation, in accordance with an exemplary embodiment of the present invention;

FIG. 8 is a simplified pictorial illustration depicting all-to-all submatrix distribution, in accordance with another exemplary embodiment of the present invention; and

FIG. 9 is a simplified pictorial illustration depicting transposition of a sub-block, in accordance with exemplary embodiments of the present invention.

DETAILED DESCRIPTION OF AN EMBODIMENT

The all-to-all operation, defined in communication standards such as the Message Passing Interface (MPI) (Forum, 2015), is a collective data operation in which each process sends data to every other process in the collective group, and receives the same amount of data from each process in the group. The data sent to each process is of the same length, a, and is unique, originating from distinct memory locations. In communications standards such as MPI, the concept of operations on processes is decoupled from any particular hardware infrastructure. A collective group, as discussed herein, refers to a group of processes over which a (collective) operation is defined. In the MPI specification a collective group is called a “communicator”, while in OpenSHMEM (see, for example, www.openshmem.org/site/) a collective group is called a “team”.

Reference is now made to FIG. 1A, which is a simplified pictorial illustration of an exemplary computer system, constructed and operative in accordance with an exemplary embodiment of the present invention. The system of FIG. 1A, generally designated 110, comprises a plurality of processes 120, with data (typically data blocks) 130 flowing therebetween. The term “data blocks” (in various grammatical forms) as used herein refers to data the data sent from a member (process, rank, . . . ) i to a member j within a collective group. It is appreciated, as also explained elsewhere herein, For alltoall the size of all blocks is the same (and can be 0), while for alltoallv/w no uniformity is the size of the data blocks is assumed, and some/all of the blocks may be 0.

Exemplary methods of operation of the system of FIG. 1A are described below. In FIG. 1A, by way of non-limiting example, a plurality of CPUs (comprising CPU 1, CPU 2, and CPU N), interconnected in a system-on-chip are shown running the plurality of processes 120. Other system examples, by way of non-limiting example, include: a single CPU; a plurality of systems or servers connected by a network; or any other appropriate system. As described above, the concept of operations on processes, as described herein, is decoupled from any particular hardware infrastructure, although it is appreciated that in any actual implementation some hardware infrastructure (as shown in FIG. 1A or otherwise as described above) would be used.

Reference is now made to FIG. 1B, which is a simplified pictorial illustration of an exemplary data block layout 175 comprising a plurality of data blocks 180, and to FIG. 2, which is a simplified pictorial illustration of another exemplary data block layout 210 comprising a plurality of data blocks 220. FIG. 1B shows the exemplary data block layout 175 before an all-to-all operation is applied, while FIG. 2 shows a corresponding data bock layout 210 after the all-to-all operation is applied. Each data block 180 in FIG. 1B and each data block 220 in FIG. 2 corresponds to a vector of length a.

The algorithms used to implement the all-to-all algorithm tend to fall into two categories—direct exchange and aggregation algorithms.

All-to-all aggregation algorithms are aimed at reducing latency costs, which dominate short data transfers. The all-to-all aggregation algorithms employ data forwarding approaches, to cut down on the number of messages sent, thus reducing latency costs. Such approaches gather/scatter the data from/to multiple sources, producing fewer larger data transfers, but send a given piece of data multiple times. As the number of communication contexts participating in the collective operation becomes too large, aggregation techniques become less efficient than a direct data exchange; this is due to the growing cost of transferring a given piece of data multiple times. The all-to-all algorithms take advantage of the fact that the data length a is a constant of the algorithm, providing sufficient global knowledge to coordinate data exchange through intermediate processes.

The direct exchange algorithms are typically used for all-to-all instances where the length of data being transferred, a, is above a threshold where bandwidth contributions dominate, or when the aggregation techniques aggregate data from too many processes, causing the aggregation techniques to be inefficient.

With growing system sizes, the need to support efficient implementations of small data all-to-all exchanges is increasing, as this is a data exchange pattern used by many High-Performance Computing (HPC) applications. The present invention, in exemplary embodiments thereof, presents a new all-to-all algorithm designed to increase the efficiency of small data exchanges over the full range of communicator sizes. This includes a new aggregation-based algorithm suitable for small data individualized all-to-all data exchange and may be viewed as transposing a distributed matrix. While reference to transposing, in various grammatical forms, are used throughout the present specification and claims, it is appreciated that transposing comprises a way to conceptualize algorithms in accordance with exemplary embodiments of the present invention; for example and without limiting the generality of the foregoing statement, there may be no such conceptualization at the level of (for example) the MPI standard. Such transposing comprises, in exemplary embodiments, changing the position of blocks relative to other blocks, without changing the structure within any block. The algorithms described herein with reference to exemplary embodiments of the present invention benefit from the large amount of concurrency available in the network and is designed to be simple and efficient for implementation by network hardware. Both switching hardware and Host-Channel-Adapter implementations are, in exemplary embodiments, targeted by this new design.

The individualized all-to-all-v/w algorithm is in certain respects similar to the individualized all-to-all data exchange. The individualized all-to-all-w algorithm differs from the all-to-all-v algorithm, in that the data type of each individual transfer may be unique across the function. A change is made to the all-to-all algorithm to support this collective operation. More specifically regarding data type: data being transferred using the MPI standard's interface specified a data type for all data, such as MPI_DOUBLE for a double precision word. The alltoallv interface specifies that all data elements are of the same data type. Alltoallw allows a different data type to be specified for each block of data, such as, for example, specifying a data type for data going from process i to process j.

The all-to-all-v/w operation is used for each process to exchange unique data with every other process in the group of processes participating in this collective operation. The size of data exchanged between two given processes may be asymmetric, and each pair may have a different data pattern than other pairs, with large variations in the data sizes being exchanged. A given rank need only have local, API-level information on the data exchanges in which it participates.

The individualized all-to-all-v/w algorithm aimed at the hardware implementation is somewhat similar to the individualized all-to-all algorithm, but requires more meta-data describing the detailed data lengths for implementation. In addition, only messages below a prespecified threshold are handled with this algorithm. A direct data exchange is used for the larger messages.

Previously, the algorithms for all-to-all function implementation have fallen into two broad categories:

• • Direct data exchange
  • Aggregation algorithms

The base algorithm definition describes data exchange between all pairs of processes in the collective group, or MPI communicator in the MPI definition. The term “base algorithm” refers to an algorithm definition at the interface level—logically what the function is/does, not how the function result is accomplished. Thus, by way of particular non-limiting example, the base description for alltoallv would be each process sending a block of data to all processes in the group. In certain exemplary embodiments of the present invention, by way of particular non-limiting example, methods are described for carrying out particular functions by aggregating data and by using communication patterns described herein. In general, the algorithm definition conceptually requires O(N²) data exchanges, where N is the group size.

Reference is now made to FIG. 3, which is a simplified pictorial illustration depicting all-to-all-v initial and final stages.

FIG. 3 provides an example of an individualized all-to-all-v, showing initial (reference numeral 310) and final (reference numeral 320) stages. In FIG. 3 notation (i,j) indicate data segment that started on rank i at position j and should be transferred to rank j at position i. Data sizes of all segments may vary (and even be zero length). Offset of send and receive locations may also vary.

The direct data exchange implementation of the function is the simplest implementation of the all-to-all function. A naïve implementation puts many messages on the network and has the potential to severely degrade network utilization by causing congestion and end-point n→1 contention. (The term “end-point”, as used herein, denotes an entity, such as a process or thread, which contributes data to a collective operation). As a result, algorithms that implement the direct data exchange use a communication pattern, such as pair-wise exchange, as shown in FIG. 4, (Jelena Pjevsivac-Grbovic, 2007), to reduce network load and end-point contention. For large message exchange, which is bandwidth limited, direct data exchange algorithms tend to make good use of network resources. However, when data exchanges are small, latency and message rate costs dominate overall algorithm costs, scale linearly with the N, and do not make good use of system resources. Specifically, FIG. 4 depicts a non-limiting example of direct pair-wise exchange pattern for exchanges involving process 0. The length of each exchange is a, with a bi-directional data exchange.

Aggregation algorithms (Ana Gainaru, 2016) have been used to implement the small data aggregation, with the Bruck (J. Bruck, 1997) algorithm being perhaps the most well-known algorithm in this class. The number of data exchanges in which each process is involved using this approach is O((k−1)*log_k(N)), where N is the collective groups size and k is the algorithm radix. FIG. 5 shows the communication pattern for one possible aggregation pattern. Specifically, FIG. 5 depicts a non-limiting example of an aggregation algorithm sending a side data pattern for arbitrary radix k, assuming N is an integer power of the algorithm radix k. N is the size of the collective group. The aggregation algorithms provides better scalability characteristics than the direct exchange. The reduction in the number of messages reduces the latency and message rate costs of the all-to-all operation but increases bandwidth related costs. If the group does not get too large, the aggregation algorithms outperform direct-exchange algorithms. The message size of each data exchange in the aggregation algorithms scales as O(a*N/k), where a is the all-to-all function message size. As a result, when groups become large, aggregation algorithms are ineffective at reducing the latency of the all-to-all data exchange and will cause the latency to exceed that of the direct data exchange algorithms.

In exemplary embodiments of the present invention, the all-to-all and all-to-all-v/w algorithm is aimed at optimizing the small data exchange by:

• • 1. Defining multiple aggregation points in the network, either switches or Host Channel Adapters (HCAs).
  • 2. Assigning aggregation points for data from sub-blocks of the processes destined to the same or other sub-blocks of the processes to individual aggregators within the network infrastructure. This data may be viewed as a sub-matrix of the distributed matrix. A single aggregator may handle multiple blocks of the submatrix from a single individualized all-to-all or all-to-all-v/w algorithm.
  • 3. The sub-blocks may be formed by discontinuous groups of processes, which are in certain exemplary embodiments formed on-the-fly to handle load imbalance in the calling application. In such a case, the matrix sub-blocks may be non-contiguous.
  • 4. The term “aggregator” is used herein to refer to an entity which aggregates a sub-matrix, transposes the same, and then says results to their final destination. In certain exemplary embodiments of the present invention, the aggregator is a logic block within an HCA. Then the present step 4 may comprise having the aggregator:
    • a. Gather data from all the sources
    • b. Shuffle the data to prepare so that data destined to a specific process may be sent in as a single message to this destination. The term “shuffle” in the present context refers to re-ordering incoming data from different source processes, such that data destined to a given destination can be conveniently handled. In certain exemplary embodiments of the present invention, data to a single destination may be copied to one contiguous memory block.
    • c. Send the data to the destinations
  • 5. Data discontinuity and the data source and/or destination is, in certain preferred embodiments, handled at the network edge, so that the aggregators handle only contiguous packed data. That is, data sent from a user or received by a user does not need to be contiguous in the user's virtual memory space; this situation can be conceived like faces of a cube, in which 2 of the 6 faces will not be in contiguous memory addresses. Hardware sends streams of contiguous data. Handling the non-contiguous→contiguous “packing” is done at the first step (either by using the CPU to pack the data into a contiguous buffer, or by using HCA gather capabilities). Similarly, unpacking non-contiguous data into user buffers can be done either by the HCA delivering the data to a contiguous destination buffer, and then using the CPU to unpack, or by using the HCA scatter capabilities. Thus, algorithmic data manipulation in the intermediate steps may treat contiguous, packed data.

The present invention, in exemplary embodiments thereof, may be viewed as using aggregation points within the network to collect data from a non-contiguous portion of a distributed matrix, transpose the data, and send the data to their destinations.

In exemplary embodiments, the invention may be summarized as follows:

• 1. Data layout is viewed as a distributed matrix, with each process holding a block of data destined to each other process. For the all-to-all algorithm, the data block size is the same for all source data blocks; while for all-to-all-v/w, the data block sizes may be of different lengths, including length zero. In notation used herein, the horizontal index indicates the data source and the vertical index its destination.
• 2. The collective operation performs a transpose of the data blocks.
• 3. To transpose the distributed matrix, the matrix is subdivided into rectangular submatrices, of dimension d_h×d_v, where d_h is the size in the horizontal dimension and d_v is the size in the vertical dimension. Subblocks need not be logically contiguous. The submatrices may be predefined, or may be determined at run-time based on some criteria, such as, by way of non-limiting example, an order of entry into the all-to-all operation.
• 4. Provide a data repacking unit, which accepts data from a specified set of sources, the data being destined to a specified set of destinations, repacks the data by destination, and sends the data to the specified destinations. In exemplary embodiments, the data repacking unit has subunits for each operation described. In certain exemplary embodiments of the present invention, an aggregator, as described herein, would comprise or make use of a data repacking unit.
• 5. The transposition of a submatrix is assigned to a given data repacking unit, with each unit being assigned multiple submatrices to transpose. In certain exemplary embodiments of the present invention, the assignment may be carried out by the control tree mentioned below in point 7; alternatively another component (such as, by way of non-limiting example, a software component) may be provided to carry out the assignment.
• 6. The data repacking unit may be implemented as appropriate within the system. For example, it may be implemented in a switch ASIC, Host-Channel-Adapter (HCA) unit, CPU or other appropriate hardware, and may be implemented in hardware, firmware, software, or in any appropriate combination thereof.
• 7. A reduction tree is used as a control tree to control the collective operation, in the following manner:
  • 7.1. Each process in the group registers itself with the control tree, by passing up an arrival notification to the control tree.
  • 7.2. Once all members of a subgroup arrive, the subgroup is marked as ready for operation (ready for send/receive).
  • 7.3. As source and destination group of a given submatrices are ready, the relevant data repacking unit schedules the data movement.
  • 7.4. Data is transferred from the source processes to the data repacking unit. This unit repacks the data and sends it to the appropriate destinations.
  • 7.5. Each source process is notified of completion, as is each destination process. In certain exemplary embodiments of the present invention, this is accomplished by the aggregator notifying the source and destination blocks of completion; by way of particular non-limiting example, this may be accomplished using the control tree.
  • 7.6. The operation completes locally at each process, once all expected data has been received, and transferring all source data is complete.In exemplary embodiments, a more detailed explanation is as follows:

In all-to-all and all-to-all-v/w algorithms, each process has a unique block of data destined for each other process in the group. The primary way all-to-all differs from all-to-all-v is in the data layout pattern. All-to-all data blocks are all of the same size, whereas the all-to-all-v/w algorithms support data blocks of differing sizes, and the data blocks need not be ordered in a monotonically increasing order within the user buffer.

The layout of blocks of data for the all-to-all algorithm may be viewed as a distributed matrix, with the all-to-all algorithm transposing this block distribution. It is important to note that, in exemplary embodiments of the present invention, the data within each block is not rearranged in the transposition, just the order of the data blocks themselves.

FIG. 6 shows an exemplary all-to-all data source data block layout for a group of size six, showing an exemplary initial distribution for an all-to-all operation. Each column represents that data block each process has for all the other processes. Each block is labeled with a two index label, with the first index indicating the process from which the data originates with the second index being the rank (the term “rank”, being used in accordance with the MPI standard, in which each member of the communicator (the group of processes that defines the collective) is given a rank, or an ID) of that block's destination process.

After the all-to-all operation is applied to the data in the example of FIG. 6, with the data blocks transposed, the data-block layout displayed in FIG. 7 results.

With the all-to-all-v/w algorithms a similar data transposition is performed. Such transform differs as follows:

• • 1. The data size may differ between blocks, and may even be of length zero.
  • 2. The blocks of data both at the source and destination buffers are not necessarily arranged in an increasing order by destination (source buffer) or source (result buffer). The actual block ordering is specified as part of the all-to-all-v/w operations.Therefore, similar communication patterns may be used to implement all-to-all and all-to-all-v/w.

The actual matrix transform is performed over sub-blocks of data. The term “the “actual matrix transform” is used herein because the blocks of data transfer defined by the operation can be viewed as a matrix transform, when each element in the matrix is a block of data. The columns of the matrix are the blocks of data owned by each process. Each process has a block of data associated with every process in the group, so the matrix can be viewed as a square matrix. For alltoall, the size of all the blocks is identical, for alltoall-v and alltoall-w, block sizes may be different. From a block-like view of the data layout (not the actual size of each block) alltoall-v and alltoall-w still are square.

For the purpose of the transform, horizontal submatrix dimension, d_h, and vertical submatrix dimension, d_v, are defined. The sub-block dimensions need not be an integer divisor of the full matrix dimension, and d_h and d_v need not be equal. Incomplete sub-blocks are permitted; that is, for a given group size, there are subgroups for which the ratio of the groups size to the sub-block size is not an integer. This situation gives “leftover” blocks at the edges. By way of particular non-limiting example, such “leftover” blocks would be present in a matrix of size 11, with sub-blocks of size 3. Finally, the vertical and horizontal ranges of values in the full matrix need not be contiguous, e.g., when mapped onto the full matrix, such a submatrix may be distributed into several different contiguous blocks of data over the matrix.

As an example, if we take d_h=h_v=2, and we use processes group {1,2}, {0,3} and {4,5} to sub-block the matrix, FIG. 8 uses coding [a] through [i] to show how the full matrix, may, in a non-limiting example, be subdivided into 2×2 sub-blocks. Note that there are three distributed sub-blocks in the example: 1) data blocks (0,0)(0,3)(3,0)(3,3), shown as [a]; 2) data blocks (0,1)(0,2)(3,1)(3,2), shown as [c]; and 3) (1,0)(2,0)(1,3)(2,3), shown as [b].

The full end-to-end all-to-all is orchestrated, in exemplary embodiments of the present invention, using a reduction tree. As processes make a call to the collective operation, the reduction tree is used by each process to register with the collective operation. When all members of a sub-group have registered with the operation, the sub-group is marked as active. When both source and destination subgroup are active, that subgroup may be transposed.

In certain exemplary embodiments of the present invention, the collective operation is executed in the following manner:

• • 1. Each process in the group registers itself with the control tree, by passing up an arrival notification to the controller.
  • 2. Once all members of a subgroup arrive, this subgroup is marked as ready for operation.
  • 3. As source and destination groups are ready, these are paired and assigned to a data repacking unit.
  • 4. Data is transferred from the source processes to the data repacking unit. This unit repacks the data and sends it to the appropriate destinations.
  • 5. Each source process is notified of completion, as it each destination process.
  • 6. The operation completes locally at each process, once all expected data has been received, and transferring all source data is complete.

FIG. 9 show how one of the data-repacking units 910 in the system is used, in a non-limiting exemplary embodiment, to transpose the submatrix defined by the horizontal subgroup {0,3} and vertical subgroup {1,2}. Processes 0 and 3 each sends its portion of the submatrix to the data-repacking unit, which rearranges the data, sending to processes 1 and 2. In the specific non-limiting example shown in FIG. 9, process 0 has the data elements (0,1) and (0,2), process 3 has the data elements (3,1) and (3,2). This data is sent to the controller which sends (0,1) and (3,1) to process 1 and (0,2) and (3,2) to process 2. Final data-placement in the result buffer is handled by the end-point. In general, in exemplary embodiments the repacking unit 910 treats all data processed thereby as a contiguous “blob” of data—the repacking unit 910 does not recognize any structure in the data. The final data distribution at the end-points within each block may be contiguous, in which case the repacking unit and the destination process will have the same view of the data. However, the final data layout at the destination process may be non-contiguous, in which case it is the end-point that distributes the data appropriately at the destination. It is appreciated that the end-point or any other appropriate system component may distribute the data appropriately.

REFERENCES

• Ana Gainaru, R. L. Graham, Artem Polyakov, Gilad Shainer (2016). Using InfiniBand Hardware Gather-Scatter Capabilities to Optimize MPI All-to-All (Vol. Proceedings of the 23rd European MPI Users' Group Meeting). Edinburgh, United Kingdom: ACM.
• MPI Forum, (2015). Message Passing Interface. Knoxville: University of Tennessee.
• J. Bruck, Ching-Tien Ho, Shlomo Kipnis, Derrick Weathersby (1997). Efficient algorithms for all-to-all communications in multi-port message-passing systems. In IEEE Transactions on Parallel and Distributed Systems, pages 298-309.
• Jelena Pjevsivac-Grbovic, Thara Angskun, Geroge Bosilca, Graham Fagg, Edgar Gabriel, Jack Dongarra, (2007). Performance analysis of MPI collective operations. Cluster Computing.

It is appreciated that software components of the present invention may, if desired, be implemented in ROM (read only memory) form. The software components may, generally, be implemented in hardware, if desired, using conventional techniques. It is further appreciated that the software components may be instantiated, for example: as a computer program product or on a tangible medium. In some cases, it may be possible to instantiate the software components as a signal interpretable by an appropriate computer, although such an instantiation may be excluded in certain embodiments of the present invention.

It is appreciated that various features of the invention which are, for clarity, described in the contexts of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features of the invention which are, for brevity, described in the context of a single embodiment may also be provided separately or in any suitable subcombination.

It will be appreciated by persons skilled in the art that the present invention is not limited by what has been particularly shown and described hereinabove. Rather the scope of the invention is defined by the appended claims and equivalents thereof:

Claims

1. A method comprising: providing a plurality of processes, each of said plurality of processes being configured to hold a block of data destined for others of said plurality of processes;providing at least one instance of data repacking circuitry comprising: receiving circuitry configured to receive at least one block of data from at least one source process of the plurality of processes;repacking circuitry configured to repack received data in accordance with at least one destination process of the plurality of processes; andsending circuitry configured to send the repacked data to said at least one destination process of the plurality of processes;receiving a set of data for all-to-all data exchange, the set of data being configured as a matrix, the matrix being distributed among the plurality of processes;transposing the data by: each of the plurality of processes sending matrix data from said process to said data repacking circuitry; and said data repacking circuitry receiving, repacking, and sending the resulting matrix data to destination processes; andproviding a control tree configured to control the plurality of processes and the repacking circuitry,wherein the control tree is further configured to: receive registration messages from each of the plurality of processes;mark a given subgroup of the plurality of processes as ready for operation when registration messages have been received from all members of said given subgroup;when a given subgroup which is a source subgroup and a corresponding subgroup which is a destination subgroup are ready for operation, pair the given source subgroup and the given destination subgroup and assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry; andnotify each said source subgroup and each said destination subgroup when operations relating to each said source subgroup and each said destination subgroup have completed.

2. The method according to claim 1, and wherein the control tree is configured, in addition to pairing the given source subgroup and the given destination subgroup, to assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry.

3. The method according to claim 1, and also including providing assigning circuitry other than the control tree, the assigning circuitry being configured to assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry.

4. The method according to claim 1 and wherein the control tree comprises a reduction tree.

5. The method according to claim 4, and also including providing assigning circuitry other than the control tree, the assigning circuitry being configured to assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry.

6. Apparatus comprising: receiving circuitry configured to receive at least one block of data from at least one source process of a plurality of processes, each of said plurality of processes being configured to hold a block of data destined for others of said plurality of processes;at least one instance of data repacking circuitry configured to repack received data in accordance with at least one destination process of the plurality of processes; andsending circuitry configured to send the repacked data to said at least one destination process of the plurality of processes,the apparatus being configured to receive a set of data for all-to-all data exchange, the set of data being configured as a matrix, the matrix being distributed among the plurality of processes, andthe apparatus being further configured to transpose the data by: receiving, from each of the plurality of processes, matrix data from said process at the repacking circuitry; and the data repacking circuitry receiving, repacking, and sending the resulting matrix data to destination processes,the apparatus comprising a control tree configured to control the plurality of processes and the repacking circuitry, andwherein the control tree is further configured to:receive registration messages from each of the plurality of processes; mark a given subgroup of the plurality of processes as ready for operation when registration messages have been received from all members of said given subgroup;when a given subgroup which is a source subgroup and a corresponding subgroup which is a destination subgroup are ready for operation, pair the given source subgroup and the given destination subgroup and assign the given source subgroup and the given destination subgroup to an instance of data repacking circuitry; andnotify each said source subgroup and each said destination subgroup when operations relating to each said source subgroup and each said destination subgroup have completed.

7. Apparatus according to claim 6, and wherein the control tree is configured, in addition to pairing the given source subgroup and the given destination subgroup, to assign the given source subgroup and the given destination subgroup to a given instance of data repacking circuitry.

8. Apparatus according to claim 6, and also comprising assigning circuitry other than the control tree, the assigning circuitry being configured to assign the given source subgroup and the given destination subgroup to a given instance of data repacking circuitry.

9. Apparatus according to claim 8 and wherein the control tree comprises a reduction tree.

10. Apparatus according to claim 6 and wherein the control tree comprises a reduction tree.