CSC/ECE 506 Spring 2012/11a ht - Revision history

Laddaga: /* Delayed File Access Mechanism */

2012-04-23T19:26:23Z

Delayed File Access Mechanism

← Older revision		Revision as of 19:26, 23 April 2012
Line 105:		Line 105:
	<br>		<br>
	<center>[[Image:Delayedaccess.png\|700px\| Parallel file access mechanisms]]</center>		<center>[[Image:Delayedaccess.png\|700px\| Parallel file access mechanisms]]</center>
	<center> '''Figure 7. Comparison of file access mechanisms ''' </center>		<center> '''Figure 7. Comparison of file access mechanisms ''' <ref name="parallel file system">Ce-Kuen Shieh; Su Cheong Mac; Jyh-Chang Ueng, "Improving the Performance of Distributed Shared Memory Systems via Parallel File Input/Output" The Journal of Systems and Software 44(1998) 3-15 , August 1997
			[http://www.sciencedirect.com/science/article/pii/S0164121298100390 paper]</ref> </center>

	===Implementation in Cohesion system===		===Implementation in Cohesion system===

Laddaga: /* Adaptive Data Distribution Scheme */

2012-04-23T19:26:06Z

Adaptive Data Distribution Scheme

← Older revision		Revision as of 19:26, 23 April 2012
Line 95:		Line 95:
	<br>		<br>
	<center>[[Image:Metadata.png\|450px\|Representation of a data distribution scheme ]]</center>		<center>[[Image:Metadata.png\|450px\|Representation of a data distribution scheme ]]</center>
	<center> '''Figure 6. Structure of metadata head''' </center>		<center> '''Figure 6. Structure of metadata head''' <ref name="parallel file system">Ce-Kuen Shieh; Su Cheong Mac; Jyh-Chang Ueng, "Improving the Performance of Distributed Shared Memory Systems via Parallel File Input/Output" The Journal of Systems and Software 44(1998) 3-15 , August 1997
			[http://www.sciencedirect.com/science/article/pii/S0164121298100390 paper]</ref> </center>

	===Delayed File Access Mechanism===		===Delayed File Access Mechanism===

Laddaga: /* Parallel File Input/Output in Cohesion system */

2012-04-23T19:25:49Z

Parallel File Input/Output in Cohesion system

← Older revision		Revision as of 19:25, 23 April 2012
Line 88:		Line 88:
	<br>		<br>
	<center>[[Image:Design.png\|500px\| Design of the system]]</center>		<center>[[Image:Design.png\|500px\| Design of the system]]</center>
	<center> '''Figure 5. File view of parallel I/O system''' </center>		<center> '''Figure 5. File view of parallel I/O system''' <ref name="parallel file system">Ce-Kuen Shieh; Su Cheong Mac; Jyh-Chang Ueng, "Improving the Performance of Distributed Shared Memory Systems via Parallel File Input/Output" The Journal of Systems and Software 44(1998) 3-15 , August 1997
			[http://www.sciencedirect.com/science/article/pii/S0164121298100390 paper]</ref> </center>

	===Adaptive Data Distribution Scheme===		===Adaptive Data Distribution Scheme===

Laddaga: /* Quiz */

2012-04-23T19:21:55Z

Quiz

← Older revision		Revision as of 19:21, 23 April 2012
Line 175:		Line 175:

	=Quiz=		=Quiz=
	'''Question 1 : DSM are used for multi-processor systems that have more than _____ processors?'''		'''Question 1 : DSM are used for multi-processor systems that have more than _____ processors.'''
	a) 4		a) 4
	b) 8		b) 8
Line 187:		Line 187:
	d) None of the above		d) None of the above

	'''Question 3 : The Cray X1 is a _____ support for DSM?'''		'''Question 3 : The Cray X1 is a _____ support for DSM.'''
	a) Software		a) Software
	b) Hardware		b) Hardware
Line 199:		Line 199:
	d) Prefetching directives		d) Prefetching directives

	'''Question 5 : Parallel file I/O cohesion uses a?'''		'''Question 5 : Parallel file I/O cohesion uses a ___ .'''
	a) Delayed file access		a) Delayed file access
	b) Instant file access		b) Instant file access
Line 205:		Line 205:
	d) Virtual file access		d) Virtual file access

	'''Question 6 : The Cray X1 uses a?'''		'''Question 6 : The Cray X1 uses a ___ .'''
	a) Module with six processing threads		a) Module with six processing threads
	b) Module with a single processing thread		b) Module with a single processing thread
Line 211:		Line 211:
	d) Module with eight processing thread		d) Module with eight processing thread

	'''Question 7 : Hardware support is less researched due to?'''		'''Question 7 : Hardware support is less researched due to ___ .'''
	a) Cost of hardware		a) Cost of hardware
	b) Lack of software support for new hardware		b) Lack of software support for new hardware
Line 217:		Line 217:
	d) Compatibility issues		d) Compatibility issues

	'''Question 8 : The first DSM was designed in?'''		'''Question 8 : The first DSM was designed in ___ .'''
	a) 1975		a) 1975
	b) 1983		b) 1983
Line 223:		Line 223:
	d) 1990		d) 1990

	'''Question 9 : In a DSM handling messages is?'''		'''Question 9 : In a DSM handling messages is ___ .'''
	a) Relatively low cost		a) Relatively low cost
	b) Has not impact on performance		b) Has not impact on performance
Line 229:		Line 229:
	d) No cost as it is hardware supported		d) No cost as it is hardware supported

	'''Question 10 : Shasta was developed in?'''		'''Question 10 : Shasta was developed in ___ .'''
	a) 1986		a) 1986
	b) 1996		b) 1996
	c) 2006		c) 2006
	d) 1976		d) 1976


	=References=		=References=

Laddaga: /* Design */

2012-04-23T19:18:39Z

Design

← Older revision		Revision as of 19:18, 23 April 2012
Line 85:		Line 85:

	===Design===		===Design===
	The two ~~maingoals~~ in the design are to parallelize the file accesses and reduce network traffic, and to minimize modifications on the existing DSMapplications. In this parallel file I/O system, every computation node also acts as a storage node. A parallel file is partitioned into file blocks, which are scattered among the nodes so that each node has a disjoint portion of the parallel file. In the pre-run of an application, the files are placed on the root node(node 0), which is the node that initiates the DSM system and the application. The data access pattern of the application is collected and the files are then partitioned into file blocks and redistributed among all nodes via an adaptive distribution scheme. In the subsequent executions, the root node issues the read requests via a delayed file access mechanism to ensure that the file blocks are placed in the consumer nodes instead of the requester node, i.e., the root node. Whenc omputation is completed, write requests are issued by the root node via the delayed access mechanism so that the data on each node are stored in its local disk instead of the root node, thereby reducing the network traffic.		The two main goals in the design are to parallelize the file accesses and reduce network traffic, and to minimize modifications on the existing DSMapplications. In this parallel file I/O system, every computation node also acts as a storage node. A parallel file is partitioned into file blocks, which are scattered among the nodes so that each node has a disjoint portion of the parallel file. In the pre-run of an application, the files are placed on the root node(node 0), which is the node that initiates the DSM system and the application. The data access pattern of the application is collected and the files are then partitioned into file blocks and redistributed among all nodes via an adaptive distribution scheme. In the subsequent executions, the root node issues the read requests via a delayed file access mechanism to ensure that the file blocks are placed in the consumer nodes instead of the requester node, i.e., the root node. Whenc omputation is completed, write requests are issued by the root node via the delayed access mechanism so that the data on each node are stored in its local disk instead of the root node, thereby reducing the network traffic.
	<br>		<br>
	<center>[[Image:Design.png\|500px\| Design of the system]]</center>		<center>[[Image:Design.png\|500px\| Design of the system]]</center>

Laddaga at 03:39, 17 April 2012

2012-04-17T03:39:48Z

← Older revision		Revision as of 03:39, 17 April 2012
Line 1:		Line 1:
	~~=Performance of DSM Systems=~~
	=Introduction=		=Introduction=
	When dealing with a relatively small number of processors (8-16), according to Solihin 320, using a bus based shared memory structure is fine. Unfortunately, when you need to provide a shared memory structure for processors much greater than that, you will need a different set of organization. This new organization is needed due to the physical limitations of the bus. There are two ways you can create such a system. These include [http://en.wikipedia.org/wiki/Distributed_shared_memory Distributed Shared Memory] (DSM) or [http://en.wikipedia.org/wiki/Non-Uniform_Memory_Access Non-Uniform Memory Access] (NUMA). The benefits of having a DSM and NUMA is that we can now scale to a larger amount of processors. The disadvantage is that scaling in such a way may not be the most cost-effective solution, Solihin 320. For the remainder of this section, we will be discussing the performance of DSM's.		When dealing with a relatively small number of processors (8-16), according to Solihin 320, using a bus based shared memory structure is fine. Unfortunately, when you need to provide a shared memory structure for processors much greater than that, you will need a different set of organization. This new organization is needed due to the physical limitations of the bus. There are two ways you can create such a system. These include [http://en.wikipedia.org/wiki/Distributed_shared_memory Distributed Shared Memory] (DSM) or [http://en.wikipedia.org/wiki/Non-Uniform_Memory_Access Non-Uniform Memory Access] (NUMA). The benefits of having a DSM and NUMA is that we can now scale to a larger amount of processors. The disadvantage is that scaling in such a way may not be the most cost-effective solution, Solihin 320. For the remainder of this section, we will be discussing the performance of DSM's.

Laddaga: /* Optimization Implementations */

2012-04-17T03:39:32Z

Optimization Implementations

← Older revision		Revision as of 03:39, 17 April 2012
Line 48:		Line 48:
	Although a lot of research has been towards software support for DSM, there has been some research in adding some hardware support. Unfortunately, according to Shi <ref name="shi"></ref> there is a rejection of hardware support from large corporations. What will occur when using hardware support is a issue of compatibility. Fortunately, recent adoptions of certain hardware standards will allow for some hardware support on the mass level.		Although a lot of research has been towards software support for DSM, there has been some research in adding some hardware support. Unfortunately, according to Shi <ref name="shi"></ref> there is a rejection of hardware support from large corporations. What will occur when using hardware support is a issue of compatibility. Fortunately, recent adoptions of certain hardware standards will allow for some hardware support on the mass level.

			<br>
	=Optimization Implementations=		=Optimization Implementations=

Line 144:		Line 145:

	<br>		<br>

	=Glossary=		=Glossary=

Laddaga: /* Glossary */

2012-04-17T03:39:11Z

Glossary

← Older revision		Revision as of 03:39, 17 April 2012
Line 143:		Line 143:
	Understanding the X1 system, we see that the X1 node supports a cache-coherent shared memory and the Cray supports OpenMP, System V shared memory, and POSIX threads shared memory programming models <ref name="dunigan"></ref>. The Cray system also supports several DSM models. Donigan provides many benchmark graphs to show that the performance of the X1 is much greater than previous systems. Please refer to the paper for these graphs <ref name="dunigan"></ref>		Understanding the X1 system, we see that the X1 node supports a cache-coherent shared memory and the Cray supports OpenMP, System V shared memory, and POSIX threads shared memory programming models <ref name="dunigan"></ref>. The Cray system also supports several DSM models. Donigan provides many benchmark graphs to show that the performance of the X1 is much greater than previous systems. Please refer to the paper for these graphs <ref name="dunigan"></ref>

			<br>
	=Glossary=		=Glossary=

Laddaga at 03:36, 17 April 2012

2012-04-17T03:36:11Z

← Older revision		Revision as of 03:36, 17 April 2012
Line 1:		Line 1:
			=Performance of DSM Systems=
	=Introduction=		=Introduction=
	When dealing with a relatively small number of processors (8-16), according to Solihin 320, using a bus based shared memory structure is fine. Unfortunately, when you need to provide a shared memory structure for processors much greater than that, you will need a different set of organization. This new organization is needed due to the physical limitations of the bus. There are two ways you can create such a system. These include [http://en.wikipedia.org/wiki/Distributed_shared_memory Distributed Shared Memory] (DSM) or [http://en.wikipedia.org/wiki/Non-Uniform_Memory_Access Non-Uniform Memory Access] (NUMA). The benefits of having a DSM and NUMA is that we can now scale to a larger amount of processors. The disadvantage is that scaling in such a way may not be the most cost-effective solution, Solihin 320. For the remainder of this section, we will be discussing the performance of DSM's.		When dealing with a relatively small number of processors (8-16), according to Solihin 320, using a bus based shared memory structure is fine. Unfortunately, when you need to provide a shared memory structure for processors much greater than that, you will need a different set of organization. This new organization is needed due to the physical limitations of the bus. There are two ways you can create such a system. These include [http://en.wikipedia.org/wiki/Distributed_shared_memory Distributed Shared Memory] (DSM) or [http://en.wikipedia.org/wiki/Non-Uniform_Memory_Access Non-Uniform Memory Access] (NUMA). The benefits of having a DSM and NUMA is that we can now scale to a larger amount of processors. The disadvantage is that scaling in such a way may not be the most cost-effective solution, Solihin 320. For the remainder of this section, we will be discussing the performance of DSM's.

Laddaga: /* See Also */

2012-04-17T03:15:14Z

← Older revision		Revision as of 03:15, 17 April 2012
Line 240:		Line 240:


	1. More on DSM design issues, the following is a graph and the reference points to the paper it came from.[http://expertiza.csc.ncsu.edu/wiki/index.php/File:PerformanceMeasures.jpg Image]. http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=84877&isnumber=2778		1. More on DSM design issues, the following is a graph and the reference points to the paper it came from.[http://expertiza.csc.ncsu.edu/wiki/index.php/File:PerformanceMeasures.jpg Image]. <br> http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=84877&isnumber=2778

	2. http://ieeexplore.ieee.org.prox.lib.ncsu.edu/stamp/stamp.jsp?arnumber=301925		2. http://ieeexplore.ieee.org.prox.lib.ncsu.edu/stamp/stamp.jsp?arnumber=301925

	3. http://catalog.hathitrust.org/Record/008308413		3. http://catalog.hathitrust.org/Record/008308413