Transcript Distributed Systems

Course Overview
Principles of Operating Systems







Introduction
Computer System
Structures
Operating System
Structures
Processes
Process Synchronization
Deadlocks
CPU Scheduling
© 2000 Franz Kurfess








Memory Management
Virtual Memory
File Management
Security
Networking
Distributed Systems
Case Studies
Conclusions
Distributed Systems 1
Chapter Overview
Distributed Systems



Motivation
Objectives
Distributed System Structures




Distributed Communication






Task Distribution
Process Migration


Distributed File Systems




Naming and Transparency
Remote File Access
Stateful/Stateless Service
File Replication
© 2000 Franz Kurfess



Message Passing
Remote Procedure Call
Shared Memory
Distributed Coordination
Distributed Processing


Network Operating Systems
Distributed OS
Remote Services

Event Ordering
Mutual Exclusion
Atomicity
Distributed Deadlock
Important Concepts and
Terms
Chapter Summary
Distributed Systems 2
Motivation
 the
next step after connecting computers via
networks are distributed systems
 resources
are transparently available throughout the
system
 users
don’t need to be aware of the system structure
 environment
independent of the local machine
 the
location and execution of processes is
distributed
 load
balancing
 process migration
© 2000 Franz Kurfess
Distributed Systems 3
Objectives
 be
aware of benefits and potential problems of
distributed systems
 understand location independence and location
transparency
 understand mechanisms for distributing data and
computation
 know extensions of previous concepts to distributed
systems
© 2000 Franz Kurfess
Distributed Systems 4
Characteristics of Modern
Operating Systems
(review OS Structures chapter)
 microkernel architecture
 multithreading
 symmetric multiprocessing
 distributed operating systems
 object-oriented design
© 2000 Franz Kurfess
Distributed Systems 5
Distributed Operating Systems
 all
resources within the distributed system are
available to all processes if they have the right
permissions
 users
and processes don’t need to be aware of the exact
location of resources
 the
execution of tasks can be distributed over
several nodes
 transparent
to the task, user and programmer
 there
is one single file system encompassing all files
on all nodes
 transparent
© 2000 Franz Kurfess
to the user
Distributed Systems 6
Applications
Distributed
SystemsOS
Diagram
Users and
Distributed
User
Programs
Hardware
Operating
System
Server Processes
Personalities
MicroKernel
Computer
Node
© 2000 Franz Kurfess
MicroKernel
System
Services
MicroKernel
Device
Drivers
File System
MicroKernel
MicroKernel
MicroKernel
Computer Computer Computer Computer Computer
Node
Node
Node
Node
Node
Distributed Systems 7
Distributed System Structures
 Network
Operating Systems
 Distributed Operating Systems
 Remote Services
© 2000 Franz Kurfess
Distributed Systems 8
Network Operating Systems
 users
are aware of the individual machines in the
network
 resources are accessible via login or explicit transfer
of data
 remote
© 2000 Franz Kurfess
login, ftp, etc.
Distributed Systems 9
Distributed Operating Systems
 users
are unaware of the underlying machines and
networks
 in
practice, users still have knowledge about particular
machines
 remote
resources are accessible in the same way as
local resources
 practical

limitations
physical access (printer, removable media)
 movement
of data and processes is under the
control of the distributed OS
© 2000 Franz Kurfess
Distributed Systems 10
Distributed Processing
 Task
Distribution
 Data Migration
 Computation Migration
 Process Migration
© 2000 Franz Kurfess
Distributed Systems 11
Task Distribution
 allocation
of tasks to nodes
 static:
at compile or load time
 dynamical: at run time
 separation
 usually
© 2000 Franz Kurfess
of a task into subtasks
by or with the help of the programmer
Distributed Systems 12
Data Migration
 movement
of data within a distributed system
 transfer of whole files
 whenever
access to a file is requested, the whole file is
transferred to the node
 all future accesses in that session are local
 transfer
of necessary parts
 only
those parts of a file that are actually needed are
transferred
 similar to demand paging
 used in many modern systems

Sun NFS, Andrews, MS SMB protocol
© 2000 Franz Kurfess
Distributed Systems 13
Computation Migration
 the
computation (task, process, thread) is moved to
the location of the data
 large
quantities of data
 time to transfer data vs. time to execute remote
commands
 remote
a
procedure call
predefined procedure is invoked on a remote system
 message
passing
a
message with a request for an action is sent to the
remote system
 the remote system creates a process to execute the
requested task, and returns a message with the result
© 2000 Franz Kurfess
Distributed Systems 14
Process Migration
 extension
to computation migration
 a process may be executed at a site different from
the one where it was initiated
 reasons for process migration
 load
balancing
 computation speedup
 specialized hardware
 specialized software
 access to data
 OS
decides the allocation of processes
 practical
© 2000 Franz Kurfess
limitations (specialized hardware and software)
Distributed Systems 15
Remote Services
(see the chapter on processes)
 remote procedure calls (RPC)
 threads
© 2000 Franz Kurfess
Distributed Systems 16
Remote Procedure Calls
 usually
implemented on top of a message-based
communication scheme
 far less reliable than local procedure calls
 precautions
 binding
must be taken for failures
problems
 local
systems can integrate the procedure call into the
executable
 this is not possible for remote calls


fixed port number at compile time
dynamic approach (rendezvous arranged via matchmaker port)
© 2000 Franz Kurfess
Distributed Systems 17
Threads
 often
used in combination with remote procedure
calls
 threads
execute RPCs on the receiving system
 lower overhead than full processes
a
server spawns a new thread for incoming requests
 all
threads can continue concurrently
 threads don’t block each other
 example:
© 2000 Franz Kurfess
Distributed Computing Environment (DCE)
Distributed Systems 18
Distributed Computing
Environment (DCE)
 threads
package for standardizing network
functionality and protocols
 system




calls for various purposes
thread management
synchronization
condition variables
scheduling
 interoperability
 available
© 2000 Franz Kurfess
for most Unix systems, Windows NT
Distributed Systems 19
Distributed File Systems
 Naming
and Transparency
 Remote File Access
 Stateful/Stateless Service
 File Replication
 Example: Sun NFS
© 2000 Franz Kurfess
Distributed Systems 20
Distributed File System
 multi-user
file system where files may reside on
various nodes in a distributed system
 transfer time over the network as additional delay
 at
10 MBit/s, the transfer of a 1 MByte file will take about 1
second (under good conditions)
© 2000 Franz Kurfess
Distributed Systems 21
Naming and Transparency
 naming
 mapping
between logical file names as seen by the user,
and the physical location of the blocks that constitute a file
 location
transparency
 the
actual location of the file does not have to be known by
the user

a file may reside on a local system or on a central file server
 files
may be cached or replicated for performance reasons
 location
independence
 the
file may be moved, but its name doesn’t have to be
changed
© 2000 Franz Kurfess
Distributed Systems 22
Naming Schemes
 specification
of host and path
 host:local-path/file-name
 not
location transparent nor location independent
 mounting
of remote directories
 remote
directories can be attached to local directories
 can become cumbersome to maintain
 location transparent, but not location independent
 example: Sun NFS
 global
name space
 total
integration of the individual file systems
 location transparent, location independent
 example: Andrews, Sprite, Locus
© 2000 Franz Kurfess
Distributed Systems 23
Remote File Access
 remote
service
 on
top of a remote procedure call mechanism
 extension of system calls
 frequently
© 2000 Franz Kurfess
caching is used to improve performance
Distributed Systems 24
Stateful/Stateless Service
 stateful
file service
 connection
between client and server is maintained for the
duration of a session
 the server has information on the status of the client
 frequently better performance
 stateless
file service
 each
request is self-contained
 the server keeps no information on the status or previous
activities of the client
 less complex
© 2000 Franz Kurfess
Distributed Systems 25
File Replication
 several
copies of files are kept on different machines
 performance

better access times
 redundancy

loss or corruption of a file is not a big problem
 consistency

different instances must be kept identical
© 2000 Franz Kurfess
Distributed Systems 26
Sun NFS
 widely
used distributed file system
 interconnected workstations are viewed as
independent nodes with independent file systems
 files can be shared between any pair of nodes
 not
restricted to servers
 implemented
by mounting directories into local file
systems
 the
mounted directory looks like a part of the local file
system
 remote
procedure calls enable remote file
operations
© 2000 Franz Kurfess
Distributed Systems 27
NFS Diagram
Client
Server
file system calls
VFS Interface
VFS Interface
other file Unix file NFS
systems system client
RPC
Communication
Mech.
NFS Unix file other file
client system systems
RPC
Communication
Mech.
Request
Response
Operating System
Operating System
Hardware Platform
Hardware Platform
© 2000 Franz Kurfess
Distributed Systems 28
Communication in Distributed
Systems
 Message
Passing
 Remote Procedure Call
 Shared Memory
 impractical
© 2000 Franz Kurfess
for distributed systems
Distributed Systems 29
Message Passing
a
request for a service is sent from the local system
to the remote system
 the
request is sent in the form of a message
 the
receiving process accepts the message and
performs the desired service
 the
result is also returned as a message
 reliability
 acknowledgments
may be used to indicate the receipt f a
message
 synchronous
(blocking) or asynchronous (non-
blocking)
© 2000 Franz Kurfess
Distributed Systems 30
Remote Procedure Call
 often
built on top of message passing
 frequently
 parameter
implemented as synchronous calls
passing
 call
by value is much easier to implement than call by
reference
 parameter
representation
 translation
between programming languages or operating
systems may be required
© 2000 Franz Kurfess
Distributed Systems 31
RPC Diagram
Client
Application
Server
Application
local calls
local calls
Application Logic Local
(Client Side)
Stub
RPC
Communication
Mech.
Local Application Logic
Stub
(Client Side)
Request
Response
RPC
Communication
Mech.
Operating System
Operating System
Hardware Platform
Hardware Platform
© 2000 Franz Kurfess
Distributed Systems 32
Distributed Coordination
 Event
Ordering
 Mutual Exclusion
 Atomicity
 Distributed Deadlock
© 2000 Franz Kurfess
Distributed Systems 33
Distributed Coordination
 synchronization
of processes across distributed
systems
 no
common memory
 no common clock
 extension
of methods discussed in the chapters on
process synchronization and deadlocks
 not
all methods can be extended easily
© 2000 Franz Kurfess
Distributed Systems 34
Event Ordering
 straightforward
in a single system
 it
is always possible to determine if on event happens
before, at the same time, or after another event
 often expressed by the happened-before relation

defines a total ordering of the events
 timestamps
can be used in distributed systems to
determine a global ordering of events
© 2000 Franz Kurfess
Distributed Systems 35
Event Ordering Diagram
A
A
A
Time
Event
Process
© 2000 Franz Kurfess
Distributed Systems 36
Mutual Exclusion
 critical
sections which may be used by at most one
process at a time
 processes
are distributed over several nodes
 approaches
 centralized
 fully
distributed
 token passing
© 2000 Franz Kurfess
Distributed Systems 37
Centralized Approach
 one
process coordinates the entry to the critical
section
 processes wishing to enter the critical section send a
request message to the coordinator
 one process gets permission through a reply
message from the coordinator, enters the critical
section, and sends a release message to the
coordinator
 coordinator is critical
 if
it fails, a new coordinator must be determined
© 2000 Franz Kurfess
Distributed Systems 38
Fully Distributed Approach
 far
more complicated than the centralized approach
 based on event ordering with timestamps
 a process that wants to enter its critical section
sends a request (including the timestamp) to all
other processes
 it waits until it receives a reply message from all
processes before entering the critical section
 if a process is in its critical section, it won’t send a
reply message until it has left the critical section
© 2000 Franz Kurfess
Distributed Systems 39
Token Passing
 processes
are arranged in a logical ring
 one single token is passed around the distributed
system
 the holder of the token is allowed to enter the critical
section
 precautions must be taken for lost tokens
© 2000 Franz Kurfess
Distributed Systems 40
Distributed Atomicity
 atomic
a
transaction
set of operations that is either fully executed, or not at all
 in
a distributed system, the operations grouped into
one atomic transaction may be executed on different
nodes/sites
 transaction coordinator
 local
coordinator guarantees atomicity at one site
 two-phase
© 2000 Franz Kurfess
commit (2PC) protocol
Distributed Systems 41
Two-Phase Commit Protocol
 makes
sure that all sites involved either commit to a
common transaction, or abort
 Phase 1
 after
the execution of the transaction, the transaction
manager at the initiating site queries all the others if they
are willing to commit their portions of the transaction
 Phase
2
 if
all answer positively within a given time, the transaction
is committed; otherwise it must be aborted
 the outcome is reported to all sites, and they finalize the
commit or abort
© 2000 Franz Kurfess
Distributed Systems 42
Failure Handling in 2PC
 participating

site
the affected site must either redo, undo, or contact the coordinator
about the fate of the transaction
 coordinator


if a participating site has a commit (abort) on record, the
transaction must be committed (aborted)
it may be impossible to determine if and what kind of decision has
been made, and the sites must wait for the coordinator to recover
 network


for one link, it is similar to the failure of a site
for several links, the network may be partitioned
 if coordinator and all participating sites are in the same partition, the
protocol can continue
 otherwise it is similar to site failure
© 2000 Franz Kurfess
Distributed Systems 43
Distributed Deadlock
 extensions
of the methods and algorithms discussed
in the chapter on deadlocks
 deadlock prevention and avoidance
 resource
ordering
 banker’s algorithm
 prioritized preemption
 deadlock
detection
(simple case: one instance per resource type)
 centralized
 fully distributed
© 2000 Franz Kurfess
Distributed Systems 44
Deadlock Prevention
 resource-ordering
 can
be enhanced by defining a global ordering on the
resources in the distributed system
 distributed
banker’s algorithm
 one
process performs the role of the banker
 all requests for resources must go through the banker
 the banker can become the bottleneck
 prioritized
preemption
 each
process has a unique priority number
 cycles in the resource allocation graph are prevented by
preempting processes with lower priorities
© 2000 Franz Kurfess
Distributed Systems 45
Centralized Deadlock Detection
 each
site maintains a local wait-for graph
 it must be shown that the union of all the graphs
contains no cycle
 one coordinator maintains the unified graph
 time delays may lead to false cycles
 can
be avoided by using time stamps
© 2000 Franz Kurfess
Distributed Systems 46
Distributed Deadlock Detection
 partial
graphs are maintained at every site
 if a deadlock exists, it will lead to a cycle in at least
one of the partial graphs
 based on local wait-for graphs enhanced by a node
for external processes
 an
arc to that node exists if a process waits for an external
item
a
cycle involving that external node indicates the
possibility of a deadlock
 can
be verified by a distributed deadlock detection
algorithm involving message exchanges with affected sites
© 2000 Franz Kurfess
Distributed Systems 47
Important Concepts and Terms












asynchronous
atomic transactions
client/server model
communication
coordination
distributed deadlock
distributed file system
distributed operating system
event ordering
kernel
location independence
location transparency
© 2000 Franz Kurfess












message passing
microkernel
mutual exclusion
naming
network file system
network operating system
processes
remote procedure call
resources
server, services
synchronous
tasks
Distributed Systems 48
Chapter Summary
 distributed
systems extend the functionality of
computers connected through networks
 location independence and location transparency
are important aspects of distributed systems
 distribution of data and computation can achieve
better resource utilization and performance
 many aspects of distributed systems are more
complex than for local systems
© 2000 Franz Kurfess
Distributed Systems 49