New Component Hierarchy

High level components

• Execution contexts
• Communication
• Session control
• Persistence
• Security

Execution contexts

Components associated with managing execution contexts. Execution context initiation, termination, monitoring and instantiation.

• Bootstrapping
  • system (infrastructure lifecycle)
  • tool
• Execution context management
  • Monitoring
  • Responding to changes in execution context (e.g., do something when an execution context dies)
  • Initiate changes to execution contexts (e.g., kill an execution context)
  • Start execution context on an allocated resource
• Resource management (discovery, allocation, deallocation) e.g., locations (e.g., nodes, cores) or network resources
  • Discovering what resources are out there
  • Allocation step: on systems where an execution context is started upon allocation, this can return "delayed allocation" and the resources are actually allocated when the execution context is started
  • Query what was actually allocated: in the "delayed allocation" case, this can return "nothing yet".
  • Allocate for real and execute: for the "delayed allocation"

Communications

• Point-to-point transfer layer
  • Description:
    • transfer data between two end-point
  • Functionality:
    • send data (input: "VC" or "endpoint")
      • multisend data (send same message to multiple VCs)
    • receive data
• Routing layer
  • Description:
    • Translates (stream, dest) pair into "VC/endpoint"
      • If message is to be routed, VC will point to next "hop"
    • Uses the Point-to-point transfer layer to send data on each link
      • For sending to all children call pt-to-pt multisend function
    • When messages are received, route data to next hop if necessary, otherwise forward to ordering layer
• Ordering/reliability layer
  • Description:
    • Ensures messages are received correctly and in order
  • Functionality:
    • Reorder messages
    • Acks/Naks/timeouts
• Active message layer
  • Description:
    • Implement the active messaging protocol
    • Sends message through routing layer
  • Functionality:
    • Setup active messaging transfer
    • Support for receive side active message callbacks
    • Implements rendezvous protocol
• Group operations
  • Barrier synchronization
  • Allgather
  • Broadcast
• Wrappers
  • Reduction streams
    • Description:
      • Wrapper functions to implement MRNet "filter" style functionality over STCI streams
  • Point-to-point matching
    • Wrapper functions to provide matching for pt-to-pt messages
  • Quarry functions (i.e., broadcast request, and gather responses)
    • Requests: test/wait

Session Control

This component deals with managing the session. Creating new sessions, expanding/contracting sessions, merging sessions, detaching/attaching front-ends, management of streams (create, quiesce, flush, delete), topology management.

• Session manager
  • Creates session
  • Destroy session
  • Expand/contract session
  • Attach detach front-end
  • Merge session
• Stream manager
  • Description
    • Uses topology manager to locate junctions and agents
    • Connects junctions, agents and front end according to topology
    • Controls streams (quiesce, flush)
  • Junction manager
    • Description:
      • Uses file staging component to stage junctions to service locations.
      • Keeps track of where junctions binaries have been staged and should be staged
• Topology manager
  • Description
    • Keeps track of logical topology definitions ("create topology")
    • Keeps track of logical topology mappings (maps logical nodes to physical locations)
    • Keeps track of staged topologies (where junctions and agents are staged and loaded)
    • Maps logical topology
      • Associates nodes in logical topology description to physical locations
    • Stages a mapped topology
      • Uses junction manager to stage junctions and agents
      • Uses execution context component to load junctions
• File staging manager
  • Description:
    • Junction and Agent binaries are "registered" with file staging manager.
    • Agent and stream managers can use this component to stage agent and junction binaries to compute and service locations (aka nodes).
• Agent manager
  • Description:
    • Using topology manager determines where agents are located
    • Starts agents

• I/O Forwarding

• Description
  • Forwards stdout and stderr from agents to front-end
  • Forwards stdin from front-end to one or more agents
  • May compress/aggregate data or add labels to output

Persistence

Where persistent state is stored. Store anything that would need to be cleaned up. E.g., publish/subscribe, info for attaching and detaching front end, session and tool agent location.

• Data storage
  • Stores the actual data
  • Might be implemented in a distributed/replicated fashion
• Policy server
  • other components query this component regarding policy (e.g., security, recovery, error handling)
  • Uses data storage to store and retrieve policy
• Session lookup
  • Used when front-end is connecting to a session for connection info and other state associated with the session
  • Used by a session for info on merging with another session
  • Uses data storage to store and retrieve session info
• Event manager
  • Listens for events and responds according to policy (e.g., some component is unresponsive —> clean it up)

Security

This component manages credentials for access to various resources. E.g., check the credentials of a front-end requesting to connect to a session; issue credentials to a session when it's created; manage privilege level (e.g., when a service agent creates a new session agent, the appropriate privilege level must be set). Works with policy manager.

• Credential manager
  • Issue credentials
  • Verify credentials
  • Revoke credentials
• Session credential manager (one per session)
  • Validates requests for front-end connections
  • Validates requests to merge sessions
• Resource credential manager
  • Interacts with OS/environment to get authorization to create or use resources (e.g., create a process, allocate a node, ssh into a node)
• Infrastructure credential manager
  • Validates requests to connect to infrastructure
  • Validates requests to infrastructure such as to create new session or delete existing session