Loading...
Please wait, while we are loading the content...
Similar Documents
The Role of Distributed Shared Memory in Future Experimental Distributed Systems Position Statement
| Content Provider | Semantic Scholar |
|---|---|
| Author | Fleisch, Brett D. |
| Copyright Year | 2007 |
| Abstract | Parallel programming will play an important role in future experimental distributed systems. A good parallel programming environment encourages the development of parallel applications that have source code compatibility so that they can be used, tested, or developed on various machine architectures and ported easily between them. In addition, transparency in intermachine IPC is desirable so that parallel applications can be coded independent of obtrusive communication primitives. Shared memory parallel programs can support a high degree of compatibility and transparency and can be coded easily to conform to support this programming philosophy. In UNIX for example, parallel programming could be supported with 1) a uniform shared memory interface, 2) common synchronization primitives, and 3) support for Distributed Shared Memory (DSM). Another paradigm emerging in importance over the next decade will be memory resident databases . Memory resident database systems (MMDB's) store their data in physical memory and provide very high-speed access to underlying data. MMDB's are becoming an attractive paradigm as 64-bit address architectures emerge and we are able to distribute persistent data storage throughout a cluster of networked machines using DSM. MMDB's and DSM must be designed and implemented to closely cooperate to achieve correctness and performance goals. This statement will focus on applications that can benefit by using DSM in future systems. Our work focuses on how to adapt operating systems that support single site shared memory programs for a distributed environment with DSM. Although good performance can be obtained for many applications that use DSM, our work in Mirage and Mirage+ has shown that applications which use DSM oblivious to their own network configuration, may perform poorly. A significant challenge is to build DSM systems which perform well in situations where hardware support would normally absorb the overheads from applications that execute in configurations with poor processor locality or in configurations where false sharing arises. One approach to address performance concerns has led a number of researchers to suggest relaxed coherency as a primary mechanism to improve performance in DSM systems. Although some relaxed coherence approaches are not seriously objectionable, some are inappropriate for applications. We take the position that relaxed coherency DSM systems are a weak way to improve system performance, particularly when a number of other solution approaches appear promising. We are pursuing a number of mechanisms to improve performance without sacrificing strict coherence. Our performance measurements ubstantiate the argument that good performing DSM systems that use strict coherence can be designed and implemented. 1Our work has been supported, in part by, NSF-CCR-9209405 and previously by a Joint Study with IBM Corporation. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://www.cs.ucr.edu/~brett/SEDMS-IV/position.ps |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
