24 April 2002
Operating Systems

Distributed Computing
Chapter 17
------------

Collection of computers that appear to users as a single commputer

dist op sys:  make computers seem this way
network OS:   no hiding, explicit telnet/ftp, just support from the os

distributed vs. parallel

why distribute
  - economics: original mainfraims
               price/performance ratio
  - speed    : parallelize
  - inherent distribution: some apps inherently distributed
             e.g., bank, CSCW, dist. games
  - reliability: failure detection
                 reconfiguration after failure
                 recovery
               critical apps - e.g., nuclear reactor - more is better
  - incremental growth - flexibility at corporate level

Not just independent PCs - need network
  share peripherals/ data
  comunication at person to person level
  spread load over resources - CONDOR

Negatives:
  - building software is hard
      how much should be hidden from the user (transparency)
      how much should user/OS do?
  - communication network not perfect
      messageslost, overloaded, rewiring is expensive
  - data sharing - consistency problems
                 - resource allocation, deadlock
                 - secuirty

Distributed Operating System
----------------------------
migration - data
          - code
          - process/computation

  why - load balancing
      - better n/w utilization
      - natural distribution
      - resource access (h/w, s/w) + data

  Data migration:
    move/copy files - NFS (to some degree)
    replication of database
  Code migration:
    RMI stubs, Java applets
    flexibility
  Process migration:
    a.k.a. mobile agents
    move processing closer to data. e.g. database access
    issues: communication with moving processes
  Computation migration - subtle difference from process
    RPC - transparent where computation is preformed
    msg passing - explicit - synchronous or asynchronous

Transparency
  os appears to be time sharing e.g., AMOEBA
  hide distribution - SAME semantics
  Location transparency - hide location of resources
  migration transparency - resources can move without changing names
  replication transparency - OS can make copies. e.g., files
  concurrency transparency - users do not interfere with one another
  parallelism transparency - users do not see parallelization of progs
                           - handled by compiler - runtime system - os

  do you always want full transparency?  e.g., printer 

Flexibility
  how much in kernel - microkernel arguments
    IPC, mem mgmt, low level process mgmt (scheduling), low level I/O
    rest user level services
  
Reliability
  Lamport definition of a distributed system:
    "One on which I cannot get any work done because some machine I have never
     heard of has crashed" 
  Secuirty
  fault tolerance - communication faults
                  - machine faults
                  - disk crashes
  
Performance
  response timme, throughput, system utilization, network capacity consumption
  message passing overhead
  balance overhead with gain in granularity

scalability 
  don't degrade too much as add nodes
  bottlenecks - centralization points

          
-------------------
Supporting Distributed Computation
  -- both dist os and nw os

creation/destruction
  remotely create a file, move it(?), kill it

Scheduling
  explicit or transparent
  Condor, transparent

synchronization/coordination/concurrencty control
  general interaction
  semaphores - assume shared memory to store lock
  * event ordering, mutual exclusion, atomic transactions

deadlock management
  prevention
-------------------
-------------------
Models of distributed computing

Message passing/remote procedure call/distributed objects/shared memory