1. Introduction to locality sensitive approach to distributed systems
Outline
distributed system definition
distributed system model
issues specific to distributed computing
local sensitivity
local representation
clusters
spanners
using local representation

2. Distributed system (DS) definition
Distributed system is distinguished
on architectural level by coupling level
tightly coupled (parallel machines) – synchronous processors, fast and reliable communication, shared memory
loosely coupled – independent processors, (relatively) infrequent communication, limited cooperation
purpose of cooperation – provide individual users convenient and efficient access to shared resources
via single-system image - the user is supplied with a centralized view of the system: the system is composed of a single entity located in one place and dedicated to serving that particular user (the distributed nature of the system is hidden)

3. System model
message-passing model – processes do not share memory and communicate by passing messages
shared memory is not considered
point-to-point communication – direct message exchange between pairs of processes
broadcast media (wireless, busses) is not considered
p2p communication may be anonymous: process is aware of the outgoing ports of the channels but not aware of the receiving entities

4. Distributed computing issues
communication – communication costs tend to dominate the execution
incomplete knowledge
each process may not know the complete input, network topology, stage of the (global) execution
failures – due to loose coupling both the occurrence and handling of faults is specific to DS
greater incidence of individual faults
potential to fault-tolerance dues to processors’ autonomy

5. Timing, synchrony, nondeterminism
two extreme models
synchronous model – execution proceeds by pulses or cycles containing the following steps
send messages to neighbors
receive messages from neighbors
perform local computation
asynchronous model – execution is event-driven (the processes cannot consult clock): local computation and message transmission takes arbitrary long
due to arbitrary order of message delivery the execution of a distributed system is nondetermenistic – running the same algorithm with the same inputs may produce different results

6. Local-sensitivity
traditionally DS algorithms (routing, broadcast, topology update) require each process to maintain the information about the whole network
scalability problems
global knowledge is not always necessary
many tasks can be solved such that each process involves only processes in a small region around it
also, it is desirable that the cost of solving the task is proportional to the size of the region involved (not the size of the whole network)

7. Locality-preserving (LP) network representations
we consider arbitrary topologies
idea – minimize (computational and storage) costs of execution by letting each process keep an approximate view of the topology of the system
LP-representation of the topology of the system allows each process to keep enough information about the system so as to accomplish the computing task
two types of LP-representations
clustered – grouping processes in a system in connected subsets (clusters)
skeletal – maintaining the information about sparse panning subgraphs of the network

8. Using LP-representations in computing
idea – develop applications that use the LP-representations and thus minimize costs of computing
clusters – save if most of the communication in the application within a cluster
skeletal representations – ignore non-represented edges, any application can be applied directly. The solution may be less “accurate” (in terms of network representation) but also less “costly” (in terms of bandwidth)