Collective operations are building blocks for interaction patterns, that are often used in SPMD algorithms in the parallel programming context. Hence, there is an interest in efficient realizations of these operations.

A realization of the collective operations is provided by the Message Passing Interface[1] (MPI).

Definitions edit

In all asymptotic runtime functions, we denote the latency   (or startup time per message, independent of message size), the communication cost per word  , the number of processing units   and the input size per node  . In cases where we have initial messages on more than one node we assume that all local messages are of the same size. To address individual processing units we use  .

If we do not have an equal distribution, i.e. node   has a message of size  , we get an upper bound for the runtime by setting  .

A distributed memory model is assumed. The concepts are similar for the shared memory model. However, shared memory systems can provide hardware support for some operations like broadcast (§ Broadcast) for example, which allows convenient concurrent read.[2] Thus, new algorithmic possibilities can become available.

Broadcast edit

 
Information flow of Broadcast operation performed on three nodes.

The broadcast pattern[3] is used to distribute data from one processing unit to all processing units, which is often needed in SPMD parallel programs to dispense input or global values. Broadcast can be interpreted as an inverse version of the reduce pattern (§ Reduce). Initially only root   with     stores message  . During broadcast   is sent to the remaining processing units, so that eventually   is available to all processing units.

Since an implementation by means of a sequential for-loop with   iterations becomes a bottleneck, divide-and-conquer approaches are common. One possibility is to utilize a binomial tree structure with the requirement that   has to be a power of two. When a processing unit is responsible for sending   to processing units  , it sends   to processing unit   and delegates responsibility for the processing units   to it, while its own responsibility is cut down to  .

Binomial trees have a problem with long messages  . The receiving unit of   can only propagate the message to other units, after it received the whole message. In the meantime, the communication network is not utilized. Therefore pipelining on binary trees is used, where   is split into an array of   packets of size  . The packets are then broadcast one after another, so that data is distributed fast in the communication network.

Pipelined broadcast on balanced binary tree is possible in  , whereas for the non-pipelined case it takes   cost.

Reduce edit

 
Information flow of Reduce operation performed on three nodes. f is the associative operator and α is the result of the reduction.

The reduce pattern[4] is used to collect data or partial results from different processing units and to combine them into a global result by a chosen operator. Given   processing units, message   is on processing unit   initially. All   are aggregated by   and the result is eventually stored on  . The reduction operator   must be associative at least. Some algorithms require a commutative operator with a neutral element. Operators like  ,  ,   are common.

Implementation considerations are similar to broadcast (§ Broadcast). For pipelining on binary trees the message must be representable as a vector of smaller object for component-wise reduction.

Pipelined reduce on a balanced binary tree is possible in  .

All-Reduce edit

 
Information flow of All-Reduce operation performed on three nodes. f is the associative operator and α is the result of the reduction.

The all-reduce pattern[5] (also called allreduce) is used if the result of a reduce operation (§ Reduce) must be distributed to all processing units. Given   processing units, message   is on processing unit   initially. All   are aggregated by an operator   and the result is eventually stored on all  . Analog to the reduce operation, the operator   must be at least associative.

All-reduce can be interpreted as a reduce operation with a subsequent broadcast (§ Broadcast). For long messages a corresponding implementation is suitable, whereas for short messages, the latency can be reduced by using a hypercube (Hypercube (communication pattern) § All-Gather/ All-Reduce) topology, if   is a power of two. All-reduce can also be implemented with a butterfly algorithm and achieve optimal latency and bandwidth.[6]

All-reduce is possible in  , since reduce and broadcast are possible in   with pipelining on balanced binary trees. All-reduce implemented with a butterfly algorithm achieves the same asymptotic runtime.

Prefix-Sum/Scan edit

 
Information flow of Prefix-Sum/Scan operation performed on three nodes. The operator + can be any associative operator.

The prefix-sum or scan operation[7] is used to collect data or partial results from different processing units and to compute intermediate results by an operator, which are stored on those processing units. It can be seen as a generalization of the reduce operation (§ Reduce). Given   processing units, message   is on processing unit  . The operator   must be at least associative, whereas some algorithms require also a commutative operator and a neutral element. Common operators are  ,   and  . Eventually processing unit   stores the prefix sum   . In the case of the so-called exclusive prefix sum, processing unit   stores the prefix sum   . Some algorithms require to store the overall sum at each processing unit in addition to the prefix sums.

For short messages, this can be achieved with a hypercube topology if   is a power of two. For long messages, the hypercube (Hypercube (communication pattern) § Prefix sum, Prefix sum § Distributed memory: Hypercube algorithm) topology is not suitable, since all processing units are active in every step and therefore pipelining can't be used. A binary tree topology is better suited for arbitrary   and long messages (Prefix sum § Large Message Sizes: Pipelined Binary Tree).

Prefix-sum on a binary tree can be implemented with an upward and downward phase. In the upward phase reduction is performed, while the downward phase is similar to broadcast, where the prefix sums are computed by sending different data to the left and right children. With this approach pipelining is possible, because the operations are equal to reduction (§ Reduce) and broadcast (§ Broadcast).

Pipelined prefix sum on a binary tree is possible in  .

Barrier edit

The barrier[8] as a collective operation is a generalization of the concept of a barrier, that can be used in distributed computing. When a processing unit calls barrier, it waits until all other processing units have called barrier as well. Barrier is thus used to achieve global synchronization in distributed computing.

One way to implement barrier is to call all-reduce (§ All-Reduce) with an empty/ dummy operand. We know the runtime of All-reduce is  . Using a dummy operand reduces size   to a constant factor and leads to a runtime of  .

Gather edit

 
Information flow of Gather operation performed on three nodes.

The gather communication pattern[9] is used to store data from all processing units on a single processing unit. Given   processing units, message   on processing unit  . For a fixed processing unit  , we want to store the message   on  . Gather can be thought of as a reduce operation (§ Reduce) that uses the concatenation operator. This works due to the fact that concatenation is associative. By using the same binomial tree reduction algorithm we get a runtime of  . We see that the asymptotic runtime is similar to the asymptotic runtime of reduce  , but with the addition of a factor p to the term  . This additional factor is due to the message size increasing in each step as messages get concatenated. Compare this to reduce where message size is a constant for operators like  .

All-Gather edit

 
Information flow of All-Gather operation performed on three nodes.

The all-gather communication pattern[9] is used to collect data from all processing units and to store the collected data on all processing units. Given   processing units  , message   initially stored on  , we want to store the message   on each  .

It can be thought of in multiple ways. The first is as an all-reduce operation (§ All-Reduce) with concatenation as the operator, in the same way that gather can be represented by reduce. The second is as a gather-operation followed by a broadcast of the new message of size  . With this we see that all-gather in   is possible.

Scatter edit

 
Information flow of Scatter operation performed on three nodes.

The scatter communication pattern[10] is used to distribute data from one processing unit to all the processing units. It differs from broadcast, in that it does not send the same message to all processing units. Instead it splits the message and delivers one part of it to each processing unit.

Given   processing units  , a fixed processing unit   that holds the message  . We want to transport the message   onto  . The same implementation concerns as for gather (§ Gather) apply. This leads to an optimal runtime in  .

All-to-all edit

All-to-all[11] is the most general communication pattern. For  , message   is the message that is initially stored on node   and has to be delivered to node  . We can express all communication primitives that do not use operators through all-to-all. For example, broadcast of message   from node   is emulated by setting   for   and setting   empty for  .

Assuming we have a fully connected network, the best possible runtime for all-to-all is in   . This is achieved through   rounds of direct message exchange. For   power of 2, in communication round   , node   exchanges messages with node   .

If the message size is small and latency dominates the communication, a hypercube algorithm can be used to distribute the messages in time   .

 
Information flow of All-to-All operation performed on three nodes. Letters indicate nodes and numbers indicate information items.

Runtime Overview edit

This table[12] gives an overview over the best known asymptotic runtimes, assuming we have free choice of network topology.

Example topologies we want for optimal runtime are binary tree, binomial tree, hypercube.

In practice, we have to adjust to the available physical topologies, e.g. dragonfly, fat tree, grid network (references other topologies, too).

More information under Network topology.

For each operation, the optimal algorithm can depend on the input sizes  . For example, broadcast for short messages is best implemented using a binomial tree whereas for long messages a pipelined communication on a balanced binary tree is optimal.

The complexities stated in the table depend on the latency   and the communication cost per word   in addition to the number of processing units   and the input message size per node  . The # senders and # receivers columns represent the number of senders and receivers that are involved in the operation respectively. The # messages column lists the number of input messages and the Computations? column indicates if any computations are done on the messages or if the messages are just delivered without processing. Complexity gives the asymptotic runtime complexity of an optimal implementation under free choice of topology.

Name # senders # receivers # messages Computations? Complexity
Broadcast       no  
Reduce       yes  
All-reduce       yes  
Prefix sum       yes  
Barrier       no  
Gather       no  
All-Gather       no  
Scatter       no  
All-To-All       no   or  

Notes edit

  1. ^ Intercommunicator Collective Operations. The Message Passing Interface (MPI) standard, chapter 7.3.1. Mathematics and Computer Science Division, Argonne National Laboratory.
  2. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, p. 395
  3. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, pp. 396-401
  4. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, pp. 402-403
  5. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, pp. 403-404
  6. ^ Yuan, Xin (February 2009). "Bandwidth optimal all-reduce algorithms for clusters of workstations" (PDF). Journal of Parallel and Distributed Computing. 69 (2).
  7. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, pp. 404-406
  8. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, p. 408
  9. ^ a b Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, pp. 412-413
  10. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, p. 413
  11. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, pp. 413-418
  12. ^ Sanders, Mehlhorn, Dietzfelbinger, Dementiev 2019, p. 394

References edit

Sanders, Peter; Mehlhorn, Kurt; Dietzfelbinger, Martin; Dementiev, Roman (2019). Sequential and Parallel Algorithms and Data Structures - The Basic Toolbox. Springer Nature Switzerland AG. ISBN 978-3-030-25208-3.