W. C.-Y. Hsieh, "Dynamic computation migration in distributed shared memory systems," Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge, Mass., 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....level. Gobelins provides thread migration on a Linux cluster, where threads are implemented as processes. 3 III. DESIGNING MOBILE AGENT SYSTEMS Our designs are inspired by the following observations. First, computation mobility in the form of mobile agents [37] 38] 39] or thread migration [40], is essential to good performance as it exploits data locality. Second, shared variable programming is at a higher level and easier to use than message passing. Third, DSM systems as being used today suffer from false sharing (in page based DSM systems) and memory coherence protocols are not ....

....is a policy and a mechanism that decide for an agent when and where to go, and how it goes. These constitute two basic building blocks of our design. One building block is a navigation protocol, which decides when and to which destination a Messenger migrates. Such protocols exist in earlier work [40], and they follow the principle of ownercomputes [53] which for some applications is equivalent to pivot computes but for others is not. These protocols can be modified to follow pivot computes; the basic idea can be 8 a = b c x = y z for i = 1. n x = y z m = p q t = t A[i] m = p ....

[Article contains additional citation context not shown here]

W. C.-Y. Hsieh, "Dynamic computation migration in distributed shared memory systems," Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge, Mass., 1995.

....point for my work over Computation Migration because these static annotations restrict the utility and concurrency benefits of Computation Migration. Even so, my thread migration mechanism performs about an order of magnitude faster, cycle for cycle, than the Computation Migration scheme. Hsi95] describes an extension to the work where dynamic migration is implemented using a system called MCRL. Migration decisions are based on a pair of simple heuristics based on the frequency of reads and writes. Benchmarks run on the MIT Alewife system [ABC 95] indicate that computation migration ....

Wilson Cheng-Yi Hsieh. Dynamic Computation Migration in Distributed Shared Memory Systems. Ph.d. thesis, MIT, 1995.

....barriers, broadcasts, and reductions. Many shared memory applications (e.g. the SPLASH application suites [74, 84] are written in this style. Although an experimental version of CRL that supports multiple user threads per processor and migration of threads between processors is operational [31], all results reported in this thesis were obtained using the single threaded version. CRL is implemented as a library against which user programs are linked; it is written entirely in C. Both CM 5 and Alewife versions can be compiled from a single set of sources with conditionally compiled ....

....personal communication, August 1995. 84 and, if so, how frequently can be replaced with carefully selected, efficient message based primitives. The extent to which any migratory data problem can be addressed using other techniques such as function shipping [11] and computation migration [31, 32] probably also warrants further investigation. 85 Related Work CRL builds on a large body of research into the construction of distributed shared memory systems. However, as discussed in Section 1.1, four key properties distinguish CRL from other DSM systems: simplicity, portability, ....

Wilson C. Hsieh. Dynamic Computation Migration in Distributed Shared Memory Systems. PhD thesis, Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 1995.

....e ectively gives the programmer control over the size of a coherency unit. With Jackal this is under compiler and runtime control requiring no programmer intervention. CRL supports sequential memory coherency and optimizes the protocol by caching an object if it is the only user of an object. MCRL [7] is a CRL derivative that also supports dynamic computation migration. Every write operation is potentially executed at the home node of the argument to start end write, read operations are locally executed. 7. FUTURE WORK We intend to use the codemap from one run to preinsert access checks on ....

W. C. Hsieh, M. F. Kaashoek, and W. E. Weihl. Dynamic computation migration in distributed shared memory systems. In Supercomputing '96 Conference Proceedings: The international conference on high performance computing and communications, Pittsburgh, Pennsylvania, November 1996. ACM.

....concisely. However, in some areas the paper is more sparse than concise. Dynamic computation migration is brie y described in the context of thread migration and RPC, which in turn are not de ned at all. This lack of detail is somewhat of a liability. The reader must turn to Hsieh s other work [Hsi95] for complete de nitions. 3.2 Experimentation Experiments were performed on two systems, the Alewife and CM 5, using a B tree benchmark application. The signi cant di erence between the two machines is processor architecture speed versus network architecture speed: the Alewife s communication ....

.... loss Data and resource type and location A program is represented as a list of operations, each of which has the following parameters: Code size Execution time List of data and resource requirements STATIC: always migrate computation for writes, always migrate data for reads [Hsi95] REPEAT: always migrate computation for writes, migrate data for the second and all subsequent consecutive reads [Hsi95] REPEAT N: wait for n reads before a write occurs, instead of 2. This policy is intended to alleviate the penalty REPEAT su ers under a high percentage of writes. n can ....

[Article contains additional citation context not shown here]

Wilson Hsieh. Dynamic computation migration in distributed shared memory systems. TR MIT/LCS/TR-665, Massachusetts Institute of Technology, 1995.

....barriers, broadcasts, and reductions. Many shared memory applications (e.g. the SPLASH application suites [74, 84] are written in this style. Although an experimental version of CRL that supports multiple user threads per processor and migration of threads between processors is operational [31], all results reported in this thesis were obtained using the single threaded version. CRL is implemented as a library against which user programs are linked; it is written entirely in C. Both CM 5 and Alewife versions can be compiled from a single set of sources with conditionally compiled ....

....personal communication, August 1995. and, if so, how frequently can be replaced with carefully selected, efficient message based primitives. The extent to which any migratory data problem can be addressed using other techniques such as function shipping [11] and computation migration [31, 32] probably also warrants further investigation. Chapter 7 Related Work CRL builds on a large body of research into the construction of distributed shared memory systems. However, as discussed in Section 1.1, four key properties distinguish CRL from other DSM systems: simplicity, portability, ....

Wilson C. Hsieh. Dynamic Computation Migration in Distributed Shared Memory Systems. PhD thesis, Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 1995.

....kaashoek lcs.mit.edu, http: www.pdos.lcs.mit.edu kaashoek, fax: 617 258 8682. William Weihl, 130 Lytton Avenue, Palo Alto, CA 94301, weihl pa.dec.com, http: www.research.digital.com SRC staff weihl bio.html, fax: 415 853 2104. This paper is a condensed version of the first author s dissertation [10]. a b f g data Initial (a) b) c) f Processor h g f h Data migration data data Computation migration data g Figure 1. Data migration vs. computation migration. a) illustrates the initial state of the system. b) illustrates the state of the system after data migration; c) illustrates the ....

W.C. Hsieh. Dynamic Computation Migration in Distributed Shared Memory Systems. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, September 1995. Available as MIT/LCS/TR-665.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC