Distributed Systems

What is a Distributed System?

! Tanenbaum and van Renesse: A distributed system is
one that looks to its users like an ordinary, centralized,
Distributed Systems system but runs on multiple independent CPUs
! Symptoms? Shroeder:
! Multiple, independent processing units
March 5, 2001 ! Processors communicate via a hardware
interconnect
! Processing unit failures are independent
! Manage resource sharing
! State is shared among processors
2

Distributed Systems, cont. Distributed Systems, cont.

Design Issues
Naming
! Scaling ! Consistency ! What do we need to name?
! Communication ! Failures ! Processes, Services, Hosts, Objects,
! Coordination ! Security Groups, ...
! Transparency ! Heterogeneity ! Binding - Discovering and associating a name
! Naming ! Mobility to a one of these
! Load sharing ! e.g., DNS: hostname to IP address

3 4

1
 Distributed Systems, cont. Distributed Systems, cont.

Communication Consistency
! Messages can have many characteristics: ! Since we assume network links can fail at any
time, replication is required to maintain
! Length, priority, streams consistency for longer computations
! Communication medium properties affect ! Replication of data
communication performance
! Replication of computation
! bandwidth, latency, multi-cast capability,
message prioritization ! Costs associated with consistency:
! Reduction in the amount of effective
resources
! Managing extended failures
5 6

Distributed Systems, cont. Distributed Systems, cont.

Load Sharing Remote Procedure Call

! Local vs. non-local ! Introduced by Birrell and Nelson in 1985 (Xerox
PARC)
! e.g., communication failure less likely in local
clusters of processors ! Abstracts the notion of where a computation runs
! Process migration can be expensive ! Semantics can include:
! What about a system that knows the load of each ! notion that computation may fail (e.g.,
machine, then assigns computation? errors, faults, deadlocks)
! Doesn't scale well ! blocking or non-blocking
! Issues - Turning Completeness, Propagation ! synchronous vs. asynchronous
Delay ! at least once, at most once, exactly once
7 8

2
 Distributed Systems, cont. Distributed Systems, cont.

Remote Procedure Call, cont. Server Registration

! Use Discovery or Directory Service
! Implementation assumes a client/server
framework ! Each server registers a service and a version
! RPC infrastructure must implement: ! Service - typically implemented as an integer;
a set of well-known services exist
! Server Registration, Binding, Marshalling,
Message Send/Receive ! Version - a number used to select which
server to use
! Compiler/linker can implement stubs for
RPC ! Who decides which server to return to the client?
! RPC used by NFS, AFS, but not X ! Key issue: Load Balancing

9 10

Distributed Systems, cont. Distributed Systems, cont.

Marshalling End to End Design

! Assume heterogeneous computing environment ! History: Design of the Internet
! Need to have a consistent set of data types ! Q: Where should functionality be built into
! What about pointers? the Internet
! Client/Server operate in separate contexts… ! e.g., should sessions & flows be a
fundamental part of the network, or something
! What about exceptions?
implemented at higher layers?
! Exceptions can result from errors at the client
site or network problems

11 12

3
 Distributed Systems, cont.
End to End Design
! Insight - Many functions can be built at the ends
of the network (e.g., RPC). Thus, building these
functions in at lower layers will have significant
tradeoffs
! e.g., In the 70's, everyone used the DARPA-
net to log into computers. If "sessions" had
been built into the lower layers of the Internet,
web-based traffic in the 90's would not have
scaled well

13