Message driven consistency on a network of workstations
Abstract: "Message-passing and shared-memory each have their respective advantages and disadvantages in distributed parallel programming. Message-passing gives explicit control of communication but the model may require much programming effort for dynamic communication patterns and high overhea...
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Buch |
| Sprache: | Danish English |
| Veröffentlicht: |
København
1996
|
| Schriftenreihe: | Datalogisk Institut <København>: DIKU-Rapport
1996,12 |
| Schlagworte: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | Abstract: "Message-passing and shared-memory each have their respective advantages and disadvantages in distributed parallel programming. Message-passing gives explicit control of communication but the model may require much programming effort for dynamic communication patterns and high overheads when communicating complex, pointer-based data structures. With shared-memory, communication is implicit and therefore considered easier to program. But shared-memory implementations may not perform optimally for certain sharing patterns. A new hybrid programming model that integrates distributed shared memory (DSM) and message-passing is therefore proposed along with a new memory consistency model called Message-Driven Consistency (MDC). MDC is based on annotated messages that interact with the consistency mechanisms of a shared-memory The annotated messages are available to application programmers and language implementers for construction of efficient synchronization and communication patterns in distributed parallel applications. The implementation of relaxed memory consistency on distributed-memory multicomputers is frequently based on virtual memory protection mechanisms. When a virtual memory page is initially written, a copy of the page is created. Later, the copy and the page are scanned to find the new write values and the page copy is then discarded. DSM systems that implement a relaxed memory consistency model use an important data structure called a diff. This data structure contains a runlength encoding of writes to a page. It reduces network traffic and allows the implementation of protocols so multiple processors can write to a page simultaneously. This is possible because write values may later be merged We show that diffs have certain weaknesses and we propose a new memory consistency mechanism based on a more fine-grained data structure called write ranges. The write ranges allow aggressive deallocation of internal data structures that can result in significantly less network traffic and memory allocation. Two experimental systems, called CarlOS/diff and CarlOS/range, have been implemented for evaluating our proposal on a small network of high- performance workstations connected by an ATM network. The systems implement memory consistency using diffs and write ranges, respectively. The systems were initially based on the implementation of Lazy Release Consistency in the TreadMarks system. We have also implemented support for user-level fine-grained parallelism to efficiently hide remote latencies found in a distributed computing environment |
|---|---|
| Beschreibung: | X, 129 S. graph. Darst. |