Summary of 20 March 1997

• Distributed Shared Memory (cont.)
• Recovery

Distributed Shared Memory

• Dynamic Distributed Manager Scheme:
• every node does a part of management (no overloading).
• each processor either processes the request (if true owner) or forwards the request to who it thinks is the probOwner.
• chains can be collapsed.
• by setting probOwner to requestNode whenever a processor forwards a request.
• probOwner is a hint rather than the answer.
• better performance.
• very often probOwner = true owner.

• Distributed Shared Memory:
• precludes heterogeniety
• cannot have two machines of different architecture sharing memory.
• not feasible over very large distances
• very high speed network would be required.
• lot of variability in its effectiveness
• the speedup varies differently for different problems as the the number of processors is varied.
• single large address space
• this is not very desirable because the size of the virtual memory becomes very critical.

Recovery

A fault causes an error which can lead to a failure.

Failure Recovery: process of restoring an erroneous state to a valid state.

Backward Recovery Strategies:

• Operation based ( do-undo-redo logs)
• All modifications to a state are recorded to enable reversing.
• State based ( checkpointing )
• complete state of a process is saved at checkpoints.

Recovery in concurrent systems

Orphan messages	Lost Messages
received but not sent. inconsistent state. recovery algorithm required leads to Domino effect.	sent but not received. consistent state. easier algorithms exist.

Domino effect: Rolling back one process causes one or more other processes to roll back.

Strongly consistent checkpoints:

• never need to backup beyond this point.
• no messages in transit in the interval spanned by the checkpoints.
• close to stopping the system (but can be masked from user).
• less storage (history beyond this can be discarded).

Deepa Putchala