Download 14distributed.pdf

Operating Systems
CSE 410, Spring 2004
Distributed Operating Systems
Stephen Wagner
Michigan State University
Distributed Systems
• Operating Systems were designed with a single computer in mind
– Purpose of operating system was to make efficient use of
processor
– Provide users with access to shared I/O devices
• Increasingly I/O devices are shared among many machines
• Trend has been to distribute the OS across many machines
• Definition of distributed OS is fuzzy
Evolution of Distributed Systems
• Stand-alone machine
• Networked machines and FTP
• Remote Procedure Call (RPC)
• Distributed File Systems
• Process Migration
Client/Server
• Basic model for modern distributed systems
• Server:
– Controls a set of I/O devices or some data source
• Client:
– Sends requests to the server for information
• Network is a critical component
Client/Server
LAN or WAN or
Internet
server
workstation
(client)
Figure 13.1 Generic Client/Server Environment
Client/Server
Client Workstation
Presentation services
Application logic
(client portion)
Server
request
Application logic
(server portion)
response
Communications
software
protocol
interaction
Communications
software
Client
operating system
Server
operating system
Hardware platform
Hardware platform
Figure 13.2 Generic Client/Server Architecture
Classes of Client/Server Applications
• Host based processing
– client is essentially a dumb terminal
– all work is done on the server, including presentation
• Server-based processing
– client is only responsible for presentation
– all real work is done on server
– most common model
Classes of Client/Server Applications
Client
Server
Presentation logic
Application logic
Database logic
DBMS
(a) Host-based processing
Presentation logic
Application logic
Database logic
DBMS
(b) Server-based processing
Presentation logic
Application logic
Application logic
Database logic
Classes of Client/Server Applications
• Cooperative processing
– client and server share the work
• Client-based processing
– client does most of the work
– server is just a passive data source
Database logic
DBMS
(b) Server-based processing
Classes of Client/Server Applications
Presentation logic
Application logic
Application logic
Database logic
DBMS
(c) Cooperative processing
Presentation logic
Application logic
Database logic
Database logic
DBMS
(d) Client-based processing
Figure 13.5 Classes of Client/Server Applications
Client/Server: Three Tier Architecture
• Traditional Client/Server involves two tiers
• Three Tier systems include a middle tier between clients
and servers
• The middle tier provides a common gateway between
clients and a number of servers
– support old servers
– provide security
Client/Server: Three Tier Architecture
Client
Middle-tier server
(application server}
Back-end servers
(data servers)
Figure 13.6 Three-tier Client/Server Architecture
Client/Server and Caches
• Caches are used to improve performance
– used at all levels
• Recently used items will be stored in the caches
• Both the clients and servers will have caches
• Reduces network traffic
Client/Server and Caches
Network
File
Traffic
Client
Cache
Server
Traffic
Server
Cache
Disk
Traffic
Server
Disk
Server
Traffic
Client
Cache
Disk
Traffic
Client
Disk
Figure 13.7 Distributed File Cacheing in Sprite
File
Traffic
Cache Consistency
• If data is only being read, there is no problem
• If data is updated, different caches may have conflicting
data
• Challenging problem
– limiting concurrent access solves the problem, but
limits performance
Middleware
• There is a lot of diversity in the client/server model
• Standardized protocols are known as middleware
– exist “between” client and server
– simplify the development of client/server applications
– many middleware packages with a wide range of
complexity
Middleware
• Three Common Middleware Packages
– Message Passing
– Remote Procedure Calls
– Object request broker
Distributed Message Passing
Application
Application-specific
messages
Message-Oriented
Middleware
(with message queue)
Network
Transport
Server
Transport
Network
Message-Oriented
Middleware
(with message queue)
Application
Client
(a) Message-Oriented Middleware
Client
Sending
Server
process
Application-specific
procedure invocations
and returns
Network
Transport
RPC
stub
program
Application
Transport
Network
Application
Receiving
process
RPC
stub
program
(b) Remote Procedure Calls
Server
Message-passing
Transport
Network
Remote
object
requester
module
Object
request
broker
Application
ProcessId
Message
Transport
Object requests
and responses
Network
Object requests
and responses
Network
Transport
Client
Message-passing
module
Object
Server
Figure 13.12 Basic Message-Passing Primitives
(c) Object request broker
Reliability
• Message Reliability
– Is message delivery guaranteed?
– How reliable is the network?
• The simple solution is to require applications to provide reliability
– applications send acknowledgment messages
• A more sophisticated middle-ware package will provide guaranteed
message delivery
Blocking vs. Nonblocking
• In a blocking system, send and receive block the process that calls
them
– simplifies application logic
– a blocking send does not return until the message has been
delivered in a reliable service
– limits concurrency
• In a nonblocking system, a process does not have to wait for send
and receive to complete
– more flexible
– susceptible to timing errors, making application logic and
debugging harder
Remote Procedure Calls
• A more sophisticated middleware package
• Makes client server applications look like a normal
procedure call
• Client application does not need to actually know if a
function is handled by a remote procedure call or not
• Usually are synchronous, but can be asynchronous
Application
Message-Oriented
Middleware
(with message queue)
Network
Transport
Transport
Network
Message-Oriented
Middleware
(with message queue)
Application
Remote Procedure Calls
Application-specific
messages
(a) Message-Oriented Middleware
Client
Server
Application-specific
procedure invocations
and returns
Network
Transport
RPC
stub
program
Application
Transport
Network
Application
RPC
stub
program
(b) Remote Procedure Calls
Client
Server
Transport
Network
Remote
object
requester
Network
Transport
Object
request
broker
Application
Transport
Object requests
and responses
Network
Object requests
and responses
Object
Server
Remote Procedure Calls
Client
application
Local
response
Remote server
application
Local
response
Local
procedure
calls
Local application
or
operating system
Local
response
Local stub
RPC
mechanism
Local
procedure
call
Local stub
Remote procedure call
Remote procedure call
Figure 13.13 Remote Procedure Call Mechanism
RPC
mechanism
Clusters
• In the client/server model applications are not really distributed
– Most of the work is typically done on one machine
• A workstation cluster is a set of workstations (called nodes) that
are all able to work on a problem together
• Advantages of clusters
–
–
–
–
highly scalable
incremental scalability (able to add machines easily)
high availability
cost
Clusters
• There is a wide range of cluster architectures
– The individual work stations can be very independent or closely
coupled
• Single-system image architecture
– Single entry point: user logs onto the cluster, not an individual
computer
– Single file hierarchy: user sees a single file hierarchy
– Single virtual networking: any node can access any other node,
independent of how nodes are physically connected
– Single process space: a process on any node can create or
communicate with any other process on a remote node
Beowulf
• A cluster technology for PC’s running Linux
• Uses mass market commodity components, making it
inexpensive
• The PC’s are connected by a dedicated, private network
• Each node can operate as an autonomous Linux system
• Cluster can also have a single-system image
Process Migration
• In an advanced distributed OS’s, a process should be
able to migrate from one process to another
– Load balancing
– Fairness
– Failure management
• True process migration is rare