IEEE Parallel & Distributed Technology, Summer 1996
distributed memory management
Multiprocessor systems and shared memory:
Distributed shared memory (DSM) combines the two concepts.
Research Focus of DSM systems:
DSM System Structure:
States and current locations of data blocks maintained in the directory.
Research from the mid-80s.
Multiprocessor (scalability) and multicomputer (communication latency) issues converging??
3 Key Issues:
Two basic problems:
1. data is distributed to minimize access latency
2. coherent view of data must be maintained without too much overhead.
1. Replication--make copies of the data (best for read sharing)
2. Migration--move copy of the data (best when single node access at a time)
Implementation Level of the DSM Mechanism
Where is lookup of data done and an action must be taken on write to preserve coherence of shared data.
1. user-level library
2. operating system
3. programming language
Often done through virtual memory and page-fault handler. Relatively large pages can cause problems with false sharing when there is fine-grain parallelism.
Automatic replication of shared data--transparent for applications. Can handle finer-grain sharing. More for high-end machines.
``Extreme'' approaches can still involve some amount of mixed implementation.
Memory Consistency Models
Stronger Consistency Models
Sequential Consistency--all processors see the same order of reads and writes (e.g. through a single queue of accesses).
Processor Consistency--all processors must observe writes in the same order
Weaker Consistency Models
Requires memory becomes consistent only on explicit synchronization accesses. Otherwise reads/writes can be processed as received between synchronization accesses.
Release Consistency--separates acquisition and release for synchronization access. Modifications are propagated on execution of release (after an acquisition has been made).
Lazy Release Consistency--Make updates when the next relevant acquire is made.
Entry Consistency--Finer level of variable protection for better performance, but more programmer involvement.
Most, if not all work, is in the research domain--not in the commercial domain.
Work to be done in the three key areas plus some additional issues.
Tradeoffs between effort for programmer versus implementation, performance versus consistency.