Message Passing Interface 2

MPI-1

1. Fixed process model
2. point-to-point communications
3. collective operations.
4. communicators for safe library writing
5. utility routines

MPI-2

1. Dynamic process management.
2. one-sided communications
3. cooperative I/O

Dynamic Process Management

• MPI - 1 had a static or fixed number of processes
  • Cannot add nor delete processes
  • The cost of having idle processes may be large.
• Some applications favour dynamic spawning
  • run-time assessment of environment
  • serial applications with parallel modules
  • scavenger applications.

CAUTION: process initiation is expensive, hence requires careful thought.

MPI-2 Process Management

1. Parents spawn children
2. Existing MPI applications can connect
3. Formerly independent sub-applications can tear down communications and become independent.

Task Spawning - MPI_spawn();

1. It is a collective operation over the parent process’ communicator.
2. info parameter: how to start the new processes (host, architecture, work dir, path).
3. intercomm and errcodes are returned values.

Communicators

• MPI processes are identified by (group, rank) pairs.
• Communicators are either infra group, or
  • inter group : ranks refer to processes in the remote group.
• Processes in the parent and childern group each have thier own MPI_COMM_WORLD
• Send and Recv have a destination and a inter/infra communicator.
• Possible to merge processes or free parents from children by MPI_Intercomm_merge() and MPI_Comm_free().

One-sided Communications

For traditional MP, there is an implicit synchronisation - it may be delayed by asynchronous message passing.

In One-sided Communications,

• One process specifies all communication parameters.
  • Data transfer and synchronisation are separate.
• Typical operations are put, get, accumulate. MPI_put()

MPI-2 Remote Memory Access (RMA)

• Processes assign a portion (or window) of their address space that explicitly expose to RMA operations. MPI_Win()
• Two types of targets
  • Active target RMA: requires all processors that created the window to call MPI_Win_fence() before any RMA operations is guaranteed to complete.
    • One-sided communication: no process is req to post a recv.
    • Cooperative in that all processes must synchronize before any of them are guaranteed to have got/put data.
  • Passive target RMA: only requirment is that originating process places MPI_Win_Lock() and MPI_Win_Unlock() before and after the data transfer.
    • Transfer is guranteed to have completed on return from unlock.
    • Known as one-sided communication.
• Potential unsync can be avoided by locks or mutexes.

MPI-2 File Operations

• Positioning: explicit offset, shared pointer/ individual pointers
• Synch: blocking/non-blocking (async)
• Coordination: collective / non-collective
• File types:
  • is a datatype made up of elementary types; MPI_Int()
  • allows to specify non-contiguous accesses.
  • can be tiled. s.t. process writes to block of the file.

MPI-IO Usage

• 각 프로세서는 자신의 데이터를 따로 파일에 쓸수 있다
• 프로세서들은 하나의 파일로 데이터를 추가 할 수 있다. (로그파일 형식) - 한 파일 포인터 사용
• 프로세서들은 협력해서 하나의 큰 매트릭스를 파일에 쓸수 있다
  • 파일을 타일 하기 위한 파일타입 생성
  • 각가의 포인터를 이용
  • 모음 명령어를 이용하여 데이터 셔플링 가능케 함
• 병렬 파일 시스템이 사용될수 있지만 그냥 평범한 파일 시스템으로 보일 수 있음

Simple MPI I/O

각 개인 포인터를 이용한 협력 파일 명령어. 각각의 프로세서의 메모리가 할당된 파일 타일에 데이터를 넣는 형식

• MPI_File_Open() 을 이용하여 커뮤니케이터와, 각 개인 포인터와 공유 파일 포인터를 생성함
  • Info parameter 는 성능 조절, 특이 케이스 핸들링등에 대해 파일에 공유
• 각각의 읽기 쓰기는 포지셔닝을 필요로 하는데 이는
  • MPI_File_Seek; …Read() —> 개인 파일 포인터 이용
  • MPI_File_Read_at; —> 해당 오프셋에서 직접적으로 읽기
  • MPI_File_seek_shared; … read_shared() —> 공유 파일 포인터 이용. Seek_shared is collective.
  • 읽기 쓰기로 버퍼, 카운트, 데이터 형식을 specify 한다 (보통 send/recev 처럼)
  • MPI_File_close() is also collective.

MPI-2 and Beyond

• MPI-2 는 많은 기능을 넣었는데 1의 기능보다 훨씬 느려졌고, 해당 사용 회사들의 기능들은 오랫동안 미완성이었음
• MPI-3는 단방향 통신과 non-blocking collective 를 향상 시킴