Dr. David Cronk Innovative Computing Lab University of ... - It works!

More documents

Recommendations

Info

Collective Communication: Allgatherv MPI_ALLGATHERV (sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm, ierr) › Same as MPI_GATHERV, except all processes get the result › MPI_IN_PLACE may be used for sendbuf of all processors › Similar to a gatherv followed by a bcast • If there are “holes” in the receive buffer, bcast would overwrite the “holes” • displs need not be the same on each PE David Cronk 102
Allgatherv int mycount; /* inited to # local ints */ int counts[4]; /* inited to proper values */ int displs[4]; displs[0] = 0; for (I = 1; I < 4; I++) displs[I] = displs[I-1] + counts[I-1]; MPI_Allgatherv(sbuff, mycount, MPI_INT, rbuff, counts, displs, MPI_INT, MPI_COMM_WORLD); David Cronk 103
Page 1 and 2:
An Introduction to Parallel Program
Page 3 and 4:
Day 2 Course Outline (cont) Morning
Page 5 and 6:
Day 1 Morning - Lecture Course Outl
Page 7 and 8:
Parallel Processing What is paralle
Page 9 and 10:
Parallel Processing David Cronk 9
Page 11 and 12:
Parallel Processing Why parallel pr
Page 13 and 14:
Parallel Programming Outline › Pr
Page 15 and 16:
Parallel Programming Message Passin
Page 17 and 18:
Data Distribution David Cronk 17
Page 19 and 20:
Page 21 and 22:
What is MPI Introduction to MPI ›
Page 23 and 24:
What is MPI A standard library spec
Page 25 and 26:
Continuing to grow! What is MPI ›
Page 27 and 28:
Page 29 and 30:
Getting Started with MPI MPI contai
Page 31 and 32:
MPI_INIT (ierr) ierr: Integer error
Page 33 and 34:
Basic Program Structure program mai
Page 35 and 36:
MPI_COMM_RANK (comm, rank, ierr) co
Page 37 and 38:
program main include ‘mpif.h’ i
Page 39 and 40:
Message Passing Message passing is
Page 41 and 42:
Message Passing Processes are speci
Page 43 and 44:
MPI_STATUS Since a receive call onl
Page 45 and 46:
Types of data parallel MPI Programs
Page 47 and 48:
Peer/SPMD Example Compute the sum o
Page 49 and 50:
Master/Slave Example Master code: c
Page 51 and 52: Master/Slave Example Master code: c
Page 53 and 54: Unsafe Communication Patterns send
Page 55 and 56: Jacobi Iteration Initializations .
Page 57 and 58: Day 2 Course Outline (cont) Morning
Page 59 and 60: Point-to-Point Communication Modes
Page 61 and 62: MPI_ISEND (buf, cnt, dtype, dest, t
Page 63 and 64: MPI_WAIT (request, status, ierr) Re
Page 65 and 66: MPI_WAITALL (count, requests, statu
Page 67 and 68: MPI_TESTANY (count, requests, index
Page 69 and 70: Non-blocking Communication 100 cont
Page 71 and 72: Non-blocking communication J = 0 do
Page 73 and 74: Jacobi Iteration IF (myrank .EQ. 0)
Page 75 and 76: Point-to-Point Communication Modes:
Page 77 and 78: Point-to-Point Communication Modes:
Page 79 and 80: Ready Mode SLAVE MASTER while (!don
Page 81 and 82: TAKE A BREAK David Cronk 81
Page 83 and 84: Collective Communication OUTLINE
Page 85 and 86: Collective Communication Synchroniz
Page 87 and 88: Collective Communication: Bcast MPI
Page 89 and 90: Collective Communication: Gather A
Page 91 and 92: MPI_Gather for (i = 0; i < 20; i++)
Page 93 and 94: Collective Communication: Gatherv (
Page 95 and 96: Collective Communication: Gatherv (
Page 97 and 98: MPI_SCATTER IF (MYPE .EQ. ROOT) THE
Page 99 and 100: Collective Communication: Scatterv
Page 101: Collective Communication: Allgather
Page 105 and 106: Allgatherv p1 p2 p3 p4 MPI_Comm_ran
Page 107 and 108: Collective Communication: Alltoallv
Page 109 and 110: Example subroutine System_Change_in
Page 111 and 112: Example (cont) if (.not. rand_festk
Page 113 and 114: Collective Communication: Reduction
Page 115 and 116: MPI_REDUCE REAL a(n), b(n,m), c(m)
Page 117 and 118: MPI_ALLREDUCE REAL a(n), b(n,m), c(
Page 119 and 120: Collective Communication: Prefix Re
Page 121 and 122: Collective Communication: Reduction
Page 123 and 124: Collective Communication: Performan
Page 125 and 126: Derived Datatypes › Introduction
Page 127 and 128: Derived Datatypes (cont) Typemap =
Page 129 and 130: Datatype Constructors MPI_TYPE_DUP
Page 131 and 132: Datatype Constructors (cont) MPI_TY
Page 133 and 134: EXAMPLE REAL a(100,100), B(100,100)
Page 139 and 140: MPI_TYPE_CREATE_STRUCT int i; char
Page 141 and 142: MPI_TYPE_CREATE_RESIZED Struct s1 {
Page 145 and 146: Datatype Constructors: Subarrays (1
Page 147 and 148: Datatype Constructors (cont) Derive
Page 149 and 150: int i; char c[100]; MPI_Aint disp[2
Page 151 and 152: TAKE A BREAK David Cronk 151
Page 153 and 154:
Communicators and Groups Many MPI u
Page 155 and 156:
Communicators and Groups(cont) P 2
Page 157 and 158:
Group Management You can access inf
Page 159 and 160:
Course Outline (cont) Day 3 Morning
Page 161 and 162:
One Sided Communication By requirin
Page 163 and 164:
One Sided Communication Get Put Add
Page 165 and 166:
One Sided Communication Data moveme
Page 167 and 168:
Dynamic Process Management Spawning
Page 169 and 170:
MPI-I/O Introduction › What is pa
Page 171 and 172:
What is parallel I/O INTRODUCTION
Page 173 and 174:
What is parallel I/O INTRODUCTION
Page 175 and 176:
INTRODUCTION (cont) Integer dim par
Page 177 and 178:
INTRODUCTION (cont) Why use paralle
Page 179 and 180:
INTRODUCTION (cont) What is MPI-I/O
Page 181 and 182:
Recommended references MPI - The Co
show all

Dr. David Cronk Innovative Computing Lab University of ... - It works!

Create successful ePaper yourself

Delete template?

Save as template?