Applying the collective entity concept to POP-C

1
Applying the collective entity concept to POP-C

• Pierre Kuonen Christian Pérez
• EIA-Fr INRIA
• WP3 Task 3.1
• Krakow Meeting, June 27th 2006

2
Outline

• Context
• Collective entities
• POP-C
• Collective entities in POP-C
• Conclusion

3
Context

• Work started during a short visit of Christian
  Pérez to EIA-Fr (January 2006)
• Collective entity set up to specify MxN
  communications between two parallel codes
• Originally developed for distributed parallel
  objects (PaCO) and components (GridCCM)
• POP-C object oriented programming model for
  based on task programming

4
Collective entity issue
MPI
PVM
Parallel code A
Parallel code B
How to easily and efficiently handle MxN
communications?
5
Collective entity objective
What the application designer should see
HPC Entity A
HPC Entity B
and how it must be implemented!
Coll. Entity. A
Coll Entity B
6
Collective entity overview (1/3)

• Abstract model
• Foundation for building other models
• CE CORBA PaCO Distributed parallel objects
• CE CCM GridCCM Distributed parallel
  components
• Made of four models
• Definition of a distributed collective entity
• Connection and communication model
• Dynamic and parallel connections
• MxN data management model
• Support for unlimited data redistribution methods
• Exception handling model in a communication

7
Collective entity overview (2/3)

• Definition of a distributed collective entity
• May have managers

8
Collective entity overview (3/3)

• Specialization the connection managers
• Input proxy
• Enable a unique (logical) connection point
• Hide the parallel nature of the collective entity
• Output proxy
• Deals with out connection requests
• Deals with the scalar or parallel nature of
  remote entity

i
j
9
The POP Model Parallel OO Programming
Grid environment
Object
Object
execute
Object
Obj
Object
Object

• Distributed objects
• Heterogeneous
• Dynamic
• Autonomous

• Heterogeneous
• Large scale
• Unstructured
• Dynamic and unknown topology

10
POP An extension of the OO sequential model

• Generalization of the concept of (sequential)
  object
• But objects can be efficiently executed on
  different machines .nevertheless modifies as
  little as possible the OO sequential semantic
• The good proprieties of OO programming paradigm
  must be conserved
• Encapsulation, Interaction by method invocations,
  Inheritance, Polymorphism,
• Parallel object
• Various method invocation semantics
• Transparent and dynamic object allocation guided
  by the object resources need.
• Shareable, transmissible
• No explicit send/receive
• No (as few as possible) loss of performance

11
POP communication model

• The POP model uses the Inactive objects model
• Objects are active only when an method is
  invocated
• Asynchronous method invocation is possible
• The POP model defines three ways to execute a
  method
• Concurrent
• No order no atomicity
• Sequential
• Ordered and atomic regarding others sequential
  methods invocations
• Mutex
• Totally ordered and atomic

12
Method call semantics example
13
POP-C exemple
File integer.h 1 class Integer 2
public 3 Integer(int wanted, int mini) 4
Integer(char machine) 5 void Set(int
val) 6 int Get() 7 void Add(Integer
other) 8 private 9 int data 10
parclass Integer
_at_powergtwanted? mini
_at_hostmachine
seq async void Set(int val)
conc int Get()
mutex void Add(Integer other)
14
Collective entity and POP-C

• Objective
• Support parallel object implementation
• Example embed a MPI code into a POP-C object
• Definition
• A collective entity in POP-C is a collection
  (list) of objects of the same class acting
  altogether as an SPMD program
• Features
• Invoke a method on all elements (global view)
• Invoke a method on a particular element
  (individual view)

Object
Object
15
Collective method

• Rationale
• Enable a method to be collectively called on all
  elements of the collective entity
• Support for distributed parameters
• Declare with a dedicated keyword coll

b0
a0
o
ipa
opa
b1
ipb
opb
a1
b2
float x6
void foo(float x)
void bar(float x)
16
Group invocation (special case)

• Rationale
• Invoke the method on all the elements of the
  collective entity
• The parameters are replicated (broadcast mode)
• Declare with a dedicated keyword all

b0
a0
ipa
opa
o
b1
ipb
opb
a1
b2
float x6
void foo(float x)
void bar(float x)
17
Scalar method in collective entity

• Rationale
• Invoke a scalar method of a collective entity
• Define the selected element of the collective
  entity
• 4 regular cases have been selected
• first the first instance of the collection
• last the last instance of the collection
• any any instance of the collection can be
  selected
• random the instance is randomly chosen

b0
o
b1
ipb
opb
b2
18
POP-C and collective entity example

• class MySPMDclass
• public
• coll async conc void doComputation()
• first sync mutex float getResult()
• sync seq void aMethod()

gt Structural keywords (coll,first, ) are
orthogonal to behavioral keywords (sync, async,
)
19
Helper template to create collective object

• Two possibilities
• Add a new keyword to the language
• Use a template
• Second solution preferred
• template ltparclass SPMDprocgt
• parclass POPspmd
• public
• POPspmd() // n 1
• POPSpmd(int n)
• SPMDproc getSPMDprocRef(int no)
• int getSPMDsize()

20
Example of use (1/2)

• // Create a collective entity of 10 elements
• POPsmpdltMySPMDclassgt mySPMDobject(10)
• // collective call from any kind of object
• mySPMDobject.doComputation() // method is coll
• // retrieve the result
• float r mySPMDobject.GetResult() // method is
  first
• // retrieve the result from proc 4
• float r mySPMDobject.GetSPMDProcRef(4)-gtGetResul
  t()

21
Handling distributed data

• Current rationale
• Attach a distribution to parameters
• Define distribution as template
• Inherit from dataDistribution interface
• Contains methods called by the runtime to compute
  MxNcommunication as define in the collective
  entity model
• Example block cyclic for vectors
• template ltTgt class BCvectorltvectorltTgtgt public
  dataDistribution
• public
• // methods for end-user
• // methods from dataDistribution
• computeSchedule()
• private
• vectorltTgt data // example of private data to
  store
• unsigned offset // the local part of the vector

22
Conclusion

• Joint work
• Apply the collective entity model to POP-C
• Most of the rationale has been set up
• Not presented support of reduction operations
• Specializations of the CE for POP-C
• Individual access to any element of the CE
• Peaceful cohabitation of
• Regular C method,
• Scalar POP-C methods, and
• Collective methods!
• Perspective
• Finish the rationale for the TR