Subsections

8 . Processor-Aware Chare Collections

So far, we have discussed chares separately from the underlying hardware resources to which they are mapped. However, for writing lower-level libraries or optimizing application performance it is sometimes useful to create chare collections where a single chare is mapped per specified resource used in the run. The group 8 . 1 and node group constructs provide this facility by creating a collection of chares with a single chare (or branch ) on each PE (in the case of groups) or process (for node groups).


8 . 1 Group Objects

Groups have a definition syntax similar to normal chares, and they have to inherit from the system-defined class CBase _ClassName, where ClassName is the name of the group's C++ class 8 . 2 .

8 . 1 . 1 Group Definition

In the interface ( .ci ) file, we declare

 group Foo {
  // Interface specifications as for normal chares

  // For instance, the constructor ...
  entry Foo(parameters1);

  // ... and an entry method
  entry void someEntryMethod(parameters2);
};

The definition of the Foo class is given in the .h file, as follows:


 class Foo : public CBase_Foo {
  // Data and member functions as in C++
  // Entry functions as for normal chares

  public:
    Foo(parameters1);
    void someEntryMethod(parameters2);
};


8 . 1 . 2 Group Creation

Groups are created using c kNew like chares and chare arrays. Given the declarations and definitions of group Foo from above, we can create a group in the following manner:


 CProxy_Foo fooProxy = CProxy_Foo::ckNew(parameters1);

One can also use c kNew to get a CkGroupID as shown below


 CkGroupID fooGroupID = CProxy_Foo::ckNew(parameters1);

A CkGroupID is useful to specify dependence in group creations using CkEntryOptions. For example, in the following code, the creation of group GroupB on each PE depends on the creation of GroupA on that PE.


 // Create GroupA

CkGroupID groupAID = CProxy_GroupA::ckNew(parameters1);

// Create GroupB. However, for each PE, do this only 
// after GroupA has been created on it

// Specify the dependency through a `CkEntryOptions' object

CkEntryOptions opts;

opts.setGroupDepID(groupAId);

// The last argument to `ckNew' is the `CkEntryOptions' object from above

CkGroupID groupBID = CProxy_GroupB::ckNew(parameters2, opts);

Note that there can be several instances of each group type. In such a case, each instance has a unique group identifier, and its own set of branches.

8 . 1 . 3 Method Invocation on Groups

An asynchronous entry method can be invoked on a particular branch of a group through a proxy of that group. If we have a group with a proxy fooProxy and we wish to invoke entry method someEntryMethod on that branch of the group which resides on PE somePE , we would accomplish this with the following syntax:


 fooProxy[somePE].someEntryMethod(parameters);

This call is asynchronous and non-blocking; it returns immediately after sending the message. A message may be broadcast to all branches of a group (i.e., to all PEs) using the notation :


 fooProxy.anotherEntryMethod(parameters);

This invokes entry method anotherEntryMethod with the given parameters on all branches of the group. This call is also asynchronous and non-blocking, and it, too, returns immediately after sending the message.

Recall that each PE hosts a branch of every instantiated group. Sequential objects, chares and other groups can gain access to this PE-local branch using ckLocalBranch() :


 GroupType *g=groupProxy.ckLocalBranch();

This call returns a regular C++ pointer to the actual object (not a proxy) referred to by the proxy groupProxy. Once a proxy to the local branch of a group is obtained, that branch can be accessed as a regular C++ object. Its public methods can return values, and its public data is readily accessible.

Let us end with an example use-case for groups. Suppose that we have a task-parallel program in which we dynamically spawn new chares. Furthermore, assume that each one of these chares has some data to send to the mainchare. Instead of creating a separate message for each chare's data, we create a group. When a particular chare finishes its work, it reports its findings to the local branch of the group. When all the chares on a PE have finished their work, the local branch can send a single message to the main chare. This reduces the number of messages sent to the mainchare from the number of chares created to the number of processors.

For a more concrete example on how to use groups, please refer to examples/charm++/histogram_group . It presents a parallel histogramming operation in which chare array elements funnel their bin counts through a group, instead of contributing directly to a reduction across all chares.

8 . 2 NodeGroup Objects

The node group construct is similar to the group construct discussed above. Rather than having one chare per PE, a node group is a collection of chares with one chare per process , or logical node . Therefore, each logical node hosts a single branch of the node group. As with groups, node groups can be addressed via globally unique identifiers. Nonetheless, there are significant differences in the semantics of node groups as compared to groups and chare arrays. When an entry method of a node group is executed on one of its branches, it executes on some PE within the logical node. Also, multiple entry method calls can execute concurrently on a single node group branch. This makes node groups significantly different from groups and requires some care when using them.

8 . 2 . 1 NodeGroup Declaration

Node groups are defined in a similar way to groups. 8 . 3 For example, in the interface file, we declare:


  nodegroup NodeGroupType {
  // Interface specifications as for normal chares
 };

In the .h file, we define NodeGroupType as follows:


  class NodeGroupType : public CBase_NodeGroupType {
  // Data and member functions as in C++ 
  // Entry functions as for normal chares
 };

Like groups, NodeGroups are identified by a globally unique identifier of type CkGroupID . Just as with groups, this identifier is common to all branches of the NodeGroup, and can be obtained from the inherited data member thisgroup . There can be many instances corresponding to a single NodeGroup type, and each instance has a different identifier, and its own set of branches.

8 . 2 . 2 Method Invocation on NodeGroups

As with chares, chare arrays and groups, entry methods are invoked on NodeGroup branches via proxy objects. An entry method may be invoked on a particular branch of a nodegroup by specifying a logical node number argument to the square bracket operator of the proxy object. A broadcast is expressed by omitting the square bracket notation. For completeness, example syntax for these two cases is shown below:


  // Invoke `someEntryMethod' on the i-th logical node of
 // a NodeGroup whose proxy is `myNodeGroupProxy':
 myNodeGroupProxy[i].someEntryMethod(parameters);

 // Invoke `someEntryMethod' on all logical nodes of
 // a NodeGroup whose proxy is `myNodeGroupProxy':
 myNodeGroupProxy.someEntryMethod(parameters);

It is worth restating that when an entry method is invoked on a particular branch of a nodegroup, it may be executed by any PE in that logical node. Thus two invocations of a single entry method on a particular branch of a NodeGroup may be executed concurrently by two different PEs in the logical node. If this could cause data races in your program, please consult §  8.2.3 (below).


8 . 2 . 3 NodeGroups and exclusive Entry Methods

Node groups may have exclusive entry methods. The execution of an exclusive entry method invocation is mutually exclusive with those of all other exclusive entry methods invocations. That is, an exclusive entry method invocation is not executed on a logical node as long as another exclusive entry method is executing on it. More explicitly, if a method M of a nodegroup NG is marked exclusive, it means that while an instance of M is being executed by a PE within a logical node, no other PE within that logical node will execute any other exclusive methods. However, PEs in the logical node may still execute non-exclusive entry method invocations. An entry method can be marked exclusive by tagging it with the exclusive attribute, as explained in §  10.2 .

8 . 2 . 4 Accessing the Local Branch of a NodeGroup

The local branch of a NodeGroup NG, and hence its member fields and methods, can be accessed through the method NG* CProxy_NG::ckLocalBranch() of its proxy. Note that accessing data members of a NodeGroup branch in this manner is not thread-safe by default, although you may implement your own mutual exclusion schemes to ensure safety. One way to ensure safety is to use node-level locks, which are described in the Converse manual.

NodeGroups can be used in a similar way to groups so as to implement lower-level optimizations such as data sharing and message reduction.