Distributed Shared Memory for Advanced Os

7/30/2019 Distributed Shared Memory for Advanced Os

1/21

DISTRIBUTED SHARED MEMORY

By

M.Spinoza


2/21

Distributed Shared Memory

Introduction

Architecture

Advantages of DSM. Algorithms

Memory Coherence

Coherence protocols Design Issues.


3/21

Introduction

Distributed Computing

It is based on the message Passing Model in which processes

interact and share data in the form of messages.

Ex: Client-Server Model, RPC,Hoares

CSP.Distributed Shared Memory (DSM)

DSM System is a resource Management Component of a Distributed

Operating System that implements the Shared Memory model in

distributed systems, which have no physically shared memory. The shared memory model provides a virtual address space that is

shared among all nodes in a distributed system.


4/21

Architecture and Motivation

In DSM, programs access data in the

shared address space just as in traditional

virtual memory.

Each node can own data stored in the

shared address space and the ownership can

change when data moves from one node to

another.

Mapping Manager maps the shared memory address to the physical

memory. It is a layer of software implemented either in the operating

system kernel or as a runtime library routine.


5/21

Advantages of DSM

DSM Systems hides the explicit data movement and hence it iseasier to design and write parallel algorithms using DSM.

It allows complex data structures to be passed by reference thus

simplifies the development of algorithms for distributed

applications. DSM takes advantage of locality of reference exhibited by

programs and thereby cuts down on the overhead of

communicating over the network.

DSM is cheaper to build than the tightly coupled multiprocessorsystems.

It has increased processor speed and memory speed .

DSM can be easily scaled upwards.

Highly portable.


6/21

Algorithms for Implementing DSM

Issues

1. How to keep track of the location of remote data?

2. how to overcome the communication delays and highoverhead?

3. how to make shared data concurrently accessible?

Algorithms

There are 4 basic algorithms to implement DSM systems

Central Server Algorithm

Migration Algorithm

Read-Replication Algorithm

Full Replication Algorithm


7/21

Central-Server Algorithm

A central Server maintains all the shared data.

It updates the data on write requests by clients and returnsacknowledgement messages.

Duplicate write requests can be detected by

their sequence number. A failure conditions after several Retransmission

without a response.

Advantages

Simple to implementDisadvantage

Central Server can become a bottleneck

Multicasting data access requests is undesirable as it does not reduce

load at the servers.


8/21

Migration Algorithm

Data in the migration algorithm is shipped to the Location of the dataaccess request.

This algorithm allows only one node to access

a shared data at time.

The whole page or block containing the data item

migrates instead of an individual item requested.

It takes the advantage of locality of reference.

To reduce thrashing the mirage system uses a tunable parameter thatdetermines the duration for which a node can possess a shared data item.

It integrate the DSM with the virtual memory provided by OS running atindividual nodes.

To locate a data block, the algorithm makes use of server that keeps trackof the location of page.


9/21

Read-Replication Algorithm

This algorithm replicates data blocks and allows multiple nodes to haveaccess or one node to have read-write access.

The write operations is expensive.

All the copies of a shared block at various nodes

will either have to be invalidated or updated with thecurrent value to maintain the consistency of the

shared data block.

DSM must keep track of the location of all the copies of data blocks.

In the IVYsystem, the owner node of a data block keeps track of all thenodes that have a copy of the data block where as in the PLUS system, adistributed Linked-list is used for this purpose.

This algorithm has the potential to reduce the average cost of readoperations when the ratio of reads to writes is large.


10/21

Full-Replication Algorithm

This algorithm is an extension of the read-replication algorithm. It allows multiple nodes to have both read and write

access to shared data blocks.

As many nodes can write shared data concurrently,

the access to shared data must be controlled to maintain

its consistency.

consistency is maintain using gap-free sequencer. In this scheme, all nodeswishing to modify shared data will send the modifications to a sequencerassign a sequence number and multicast the modification with the sequence

number to all the nodes that have a copy of the shared data item. Each node processes the modification requests in the sequence number order.

A gap between the sequence number of a modification request and theexpected sequence number at a node indicates that one or more modificationshave been missed. This can be resolved by means of retransmission of the

modification.


11/21

Memory Coherence

A memory is coherent if the value returned by a read operation is always thevalue that the programmer expected.

There are two measures namely

1. Coherence 2. Consistency

There are several forms of memory coherenceSequential Consistency:

If the result of any execution of the operations of all the processors is thesame as if they were executed in the sequential order.

General Consistency:

If all the copies of a memory location eventually contain the same datawhen all the writes issued by every processor have completed.

Processor Consistency:

Write issued by the processor are observed in the same order in which theywere issued. Two simultaneous reads of the same location from differentprocessors may yield different results


12/21

Memory CoherenceWeak Consistency:

Before a Synchronization access can be performed, all previous regular data

accesses must be completed.

Before a regular data access can be performed, all previous synchronization

accesses must be completed.

Release Consistency:

Synchronization accesses must only be processor consistent with respect to

each other.

Synchronization operations are broken down into acquire and release. All pending acquires(e.g.: lock operation) must be done before a regular access

is done and all regular accesses must be done before a release(e.g.: unlock

operation) is done.


13/21

Coherence Protocols

To provide concurrent access, DSM systems make use of data replication,where copies of data are maintained at all the nodes accessing the data.

Two basic protocols to maintain coherence in replicas are the write-invalidate

protocol and write-update protocol.

Write-Invalidate Protocol:

A write to a shared data causes the invalidation of all copies except one before

the write can proceed.

It is suitable for applications where several updates occur between reads

Disadvantage:

Invalidations are sent to all the nodes that have copies, irrespective of whether

they will use this data or not.

Write-Update Protocol:

A write to a shared data causes all copies of that data to be updated


14/21

Coherence ProtocolsDisadvantage: It is difficult to implement, as a new value has to be sent instead of

invalidation messages.

It generates network traffic.

Cache coherence in the PLUS System:

A memory coherence manager running at each node is responsible for

maintaining the consistency.

Read Operation:

On a read fault, if the address indicates local memory, the local memory is

read. Otherwise, the local MCM sends a read request to its counterpart at thespecified remote node. The data returned by the remote MCM is passed back

to the requesting processor.

Write Operation:

Writes are always performed first on the master copy and are then propagated

to the copies linked by the copy-List.


15/21

Coherence ProtocolsUnifying Synchronization and data transfer in clouds

Reading and writing of shared data by processes is invariably controlled

by a synchronization method.

In DSM systems using data replication and write-invalidate protocol, the

reader-writer problems can be implemented using locks as follows:

(* writer process*) (*reader process*)

Loop Loop

wait(empty); wait(full);

writelock(buffer); readlock(buffer);

update buffer; read buffer;

unlock (buffer); unlock (buffer);

signal(full); signal(empty);

Endloop; Endloop;


16/21

Coherence ProtocolsType-Specific Memory Coherence in the Munin System: Following are the type-specific coherence mechanisms in Munin

Write-once objects:

These objects are replicated on demand and are accessed locally at each site.

Private objects:

These objects are accessed by single threads even though they are accessible to

all the threads of a process. It is not managed by the memory coherence system.

Write-many objects:

These objects are frequently modified by multiple threads between

synchronization points. Whenever a node updates a replicated data objects ,theupdates will be buffered.

Result objects:

These objects are not read until all parts of them are updated. Munin efficiently

maintains these objects through the delayed update mechanism.


17/21

Coherence ProtocolsSynchronization objects: Objects such as distributed locks can be employed to give threads

exclusive access to data objects.

A lock operation by a thread on the local proxy object causes the local lockserver to interact with the other lock servers to acquire the global lockassociated with the local proxy.

Migratory objects:

These objects are accessed in phases, where each phase corresponds to aseries of accesses by a single thread. Objects in critical section fall under thisclass.

Producer-Consumer objects:

These objects are typically written by one thread and read by a fixed set ofother threads.

Munin handles this type of objects by moving the objects to the site where it

will be accessed in advance. This is referred to as eager object movement.


18/21

Coherence ProtocolsRead-mostly objects:

These objects are read far more frequently than they are written.

General Read-write objects:

These objects do not exhibit particular pattern of access behavior.

Munin makes use of the Berkeley ownership protocol, which supports strictconsistency.

An object at a node can be in one of the following states:

Invalid:The object does not contain useful data.

Unowned: The object contains valid data. Other nodes may have copies of the

object. The object cannot be updated without first acquiring ownership.Owned exclusively:The object contains valid data, but it cannot be updated

before invalidating other copies.

Two types of read operations, read-shared and read-for-ownership in this

protocol.


19/21

Design Issues

The two important issues to be considered in the design of a DSM systemare as follows:

Granularity

Page Replacement.

Granularity:

It refers to the size of the shared memory unit.

By integrating DSM with the underlying memory management system, aDSM system can take the advantage of the built-in protection mechanismto detect inherent memory references.

Larger the page size ,the greater the chance for contention to access a pageby many processes.

False sharing of a page occurs when two different data items, not sharedbut accessed by two different processes, are allocated to a single page.


20/21

Design IssuesPage Replacement:

A memory management system has to address the issue of page replacement

because the size of physical memory is limited.

In DSM systems the data movement cannot be directly supported by

traditional methods such as Least Recently Used(LRU).

Data may be accessed in different modes such as shared,.. private, read-only,

writable, etc

Once a page is selected for replacement, the DSM system must ensure that

the page is not lost forever. Some systems make use of reserved memory, where each node is

responsible for certain portions of the global virtual space and reserves

memory space for those portions


21/21

Thank You

Distributed Shared Memory for Advanced Os

Documents

Transcript of Distributed Shared Memory for Advanced Os