MPI - Revision history

GethinWilliams: /* Introduction */

2010-07-26T16:35:27Z

Introduction

GethinWilliams: /* Introduction */

2010-07-26T16:34:48Z

Introduction

GethinWilliams: /* Introduction */

2010-07-26T16:34:08Z

Introduction

GethinWilliams: /* One-sided Communications */

2010-07-26T16:32:24Z

One-sided Communications

GethinWilliams: /* One-sided Communications */

2010-07-26T16:28:58Z

One-sided Communications

GethinWilliams: /* One-sided Communications */

2010-07-26T14:56:00Z

One-sided Communications

GethinWilliams: /* Load Balancing */

2010-07-26T14:49:41Z

Load Balancing

GethinWilliams: /* Load Balancing */

2010-07-26T14:48:47Z

Load Balancing

GethinWilliams: /* Exercises */

2010-07-26T14:23:39Z

Exercises

GethinWilliams at 14:22, 26 July 2010

2010-07-26T14:22:35Z

@@ Line 4: / Line 4: @@
 =Introduction=
-[[Image:Message passing1.JPG|600px|thumbnail|center]]
+[[Image:Message passing2.jpg|500px|thumbnail|center]]
 MPI-1[ref]--the first incarnation of the standard--arrived in 1994 in response to the need for a '''portable''' means to program the growing number of distributed memory computers appearing in the marketplace.  '''MPI''' stands for '''M'''essage '''P'''assing '''I'''nterface, and as its name suggests, it is an API, rather than a new programming language.  At the time of writing, MPI can be used in C, C++, Fortran-77 and Fortran-90/95 programs.  We will see that MPI-1 contained little on the topic I/O.  This was rectified in 1997 with the arrival of MPI-2[ref], which contained the MPI-IO standard (supporting parallel I/O) along with additional functionality to support the dynamic creation of processes and also one-sided communication models.

@@ Line 5: / Line 5: @@
 [[Image:Message passing1.JPG|600px|thumbnail|center]]
-[[Image:Message passing2.JPG|600px|thumbnail|center]]
+[[Image:Message passing2.jpg|600px|thumbnail|center]]
 MPI-1[ref]--the first incarnation of the standard--arrived in 1994 in response to the need for a '''portable''' means to program the growing number of distributed memory computers appearing in the marketplace.  '''MPI''' stands for '''M'''essage '''P'''assing '''I'''nterface, and as its name suggests, it is an API, rather than a new programming language.  At the time of writing, MPI can be used in C, C++, Fortran-77 and Fortran-90/95 programs.  We will see that MPI-1 contained little on the topic I/O.  This was rectified in 1997 with the arrival of MPI-2[ref], which contained the MPI-IO standard (supporting parallel I/O) along with additional functionality to support the dynamic creation of processes and also one-sided communication models.

← Older revision		Revision as of 16:34, 26 July 2010
Line 5:		Line 5:

	[[Image:Message passing1.JPG\|600px\|thumbnail\|center]]		[[Image:Message passing1.JPG\|600px\|thumbnail\|center]]
		+	[[Image:Message passing2.JPG\|600px\|thumbnail\|center]]

	MPI-1[ref]--the first incarnation of the standard--arrived in 1994 in response to the need for a '''portable''' means to program the growing number of distributed memory computers appearing in the marketplace. '''MPI''' stands for '''M'''essage '''P'''assing '''I'''nterface, and as its name suggests, it is an API, rather than a new programming language. At the time of writing, MPI can be used in C, C++, Fortran-77 and Fortran-90/95 programs. We will see that MPI-1 contained little on the topic I/O. This was rectified in 1997 with the arrival of MPI-2[ref], which contained the MPI-IO standard (supporting parallel I/O) along with additional functionality to support the dynamic creation of processes and also one-sided communication models.		MPI-1[ref]--the first incarnation of the standard--arrived in 1994 in response to the need for a '''portable''' means to program the growing number of distributed memory computers appearing in the marketplace. '''MPI''' stands for '''M'''essage '''P'''assing '''I'''nterface, and as its name suggests, it is an API, rather than a new programming language. At the time of writing, MPI can be used in C, C++, Fortran-77 and Fortran-90/95 programs. We will see that MPI-1 contained little on the topic I/O. This was rectified in 1997 with the arrival of MPI-2[ref], which contained the MPI-IO standard (supporting parallel I/O) along with additional functionality to support the dynamic creation of processes and also one-sided communication models.

← Older revision		Revision as of 16:32, 26 July 2010
Line 145:		Line 145:
	* MPI_Win_fence()		* MPI_Win_fence()
	* MPI_Accumulate(), MPI_Put & MPI_Get()		* MPI_Accumulate(), MPI_Put & MPI_Get()
		+
		+	==Exercises==
		+
		+	* Do RMA coded programs work faster on SMP machines, than over networked CE? How does the performance compare to a send/revc equivalent?

@@ Line 138: / Line 138: @@
 NB not intrinsically faster than a send and receive operation (of course that it still happening 'under-the-hood').  Rather the benefit is to provide more flexibility when it comes to implementing algorithms.
-The concept of memory windows.
+The concept of memory windows.  We can think of fence calls in terms of synchronising barriers.
 * MPI_Win_create()
 * MPI_Win_fence()
 * MPI_Accumulate(), MPI_Put & MPI_Get()

← Older revision		Revision as of 14:56, 26 July 2010
Line 135:		Line 135:

	=One-sided Communications=		=One-sided Communications=
		+
		+	NB not intrinsically faster than a send and receive operation (of course that it still happening 'under-the-hood'). Rather the benefit is to provide more flexibility when it comes to implementing algorithms.
		+
		+	The concept of memory windows.
		+
		+	* MPI_Win_create()
		+	* MPI_Win_fence()
		+	* MPI_Accumulate(), MPI_Put & MPI_Get()

← Older revision		Revision as of 14:49, 26 July 2010
Line 103:		Line 103:


−	[[Image:~~broadcast~~-tree.~~JPG~~\|600px\|thumbnail\|center]]	+	[[Image:Broadcast-tree.pdf\|600px\|thumbnail\|center]]

	How does the broadcast do better than the point-to-point?		How does the broadcast do better than the point-to-point?

← Older revision		Revision as of 14:48, 26 July 2010
Line 101:		Line 101:

	==Load Balancing==		==Load Balancing==
		+
		+
		+	[[Image:broadcast-tree.JPG\|600px\|thumbnail\|center]]

	How does the broadcast do better than the point-to-point?		How does the broadcast do better than the point-to-point?

← Older revision		Revision as of 14:23, 26 July 2010
Line 116:		Line 116:

	==Exercises==		==Exercises==
		+
		+	* Write example programs highlighting the 3 other ways of making derived types in MPI.

	=MPI-I/O=		=MPI-I/O=

@@ Line 2: / Line 2: @@
 '''MPI: Message passing for distributed memory computing'''
 [[Image:Message passing1.JPG|600px|thumbnail|center]]
 MPI-1[ref]--the first incarnation of the standard--arrived in 1994 in response to the need for a '''portable''' means to program the growing number of distributed memory computers appearing in the marketplace.  '''MPI''' stands for '''M'''essage '''P'''assing '''I'''nterface, and as its name suggests, it is an API, rather than a new programming language.  At the time of writing, MPI can be used in C, C++, Fortran-77 and Fortran-90/95 programs.  We will see that MPI-1 contained little on the topic I/O.  This was rectified in 1997 with the arrival of MPI-2[ref], which contained the MPI-IO standard (supporting parallel I/O) along with additional functionality to support the dynamic creation of processes and also one-sided communication models.
@@ Line 118: / Line 117: @@
 ==Exercises==
-=MPI-2=
+=MPI-I/O=
-=MPI-I/O=
+Introduced with MPI-2.
 Concept of views (will crop up in one-sided comms too).