Global view
    Projects
        Applications
        Scheduling server
        Database
        Project back end (interacts with scheduler via DB, files)
        web server
        Data servers
    Clients
        core client, applications, files
        single download
        preferences


Abstractions
    projects
    applications
    Workunits -> input file descriptors -> files
    Results -> output file descriptors -> files

    file descriptors can give name/FD
    files can have various properties
        files up/downloaded to servers; may be "sticky"
        client-generated scratch files; may be "sticky"
        some files originate on servers and are sent to clients;
        some files originate on client and are deleted after the WU
        some files originate on client and become input for later WUs
        some files originate on client and are uploaded (and maybe kept)

    Results are generated in advance
    (alternative: let clients fill in template)

Properties of workunits
    disk/RAM max needs
    IOPs/FLOPs/memBW needs

File transfer requests

Result sequences

Examples
    SETI@home case: generate multiple results; generate more as needed
    CP.com case: result sequences
        normal case
        interrupted/restarted case
        access to large output files

How things work on the scheduling server
    WUs and results described by XML files (has file, resource info)
    user preferences described by XML files
    file transfer requests described by XML files
    DB keeps track of
        apps
        hosts
        platforms
        users
        app versions
        core versions
        WUs
            all input files available?
        results
            unsent/sent/completed/error/timed out

Host measurements
    CPU: #cpus, int, FP, mem BW
    mem: total, cache, swap
    disk: total, free
    network: bandwidth up/down.  IP address history
    on fraction
    connected fraction

How we keep track of different hosts
    RPC seqno

The scheduling server algorithm

Project back end: examples

How things work in the client
    FSM model
    network stack
    processor slots

    account file, client state file

    main loop of client

When does the client ask for work?
    hi/lo water marks
    server backoff requests
    failed connection backoff

What are user preferences?
    max RAM to use while user active
    confirm before connect?
    don't compute while on batteries
    min time between disk I/O
    max disk space to use (absolute, or % of free)
    list of projects
        % of resources (disk, CPU)
    hours to communicate/compute
    max #bytes to upload each day
    max #bytes to download each day

How are preferences adjusted?
    Web interface
    XML file
    (extensible set of preferences w/o server change)

How a users starts up

How a user adds/removes projects

How a project starts up

Security mechanisms

Result validation and credit