MIT ISPS Software: A Space Shuttle View

• Implements non-RSVP components of the Integrated Services architecture
• Functional goals:
  • Experiments and proof of concept
  • Reference/tutorial implementation of key components
• Performance goals:
  • T1 speed / tired old machines (achieved)
  • >> ethernet speed / current machines (???)
• Freely available with no restrictions on use

Components

• ISPS Router Kernel
  • Packet classifier
  • Packet scheduler
  • Embryonic admission control
• Support tools
  • Basic configuration program
  • "netstat" extensions
  • Remote management software
• RSVP-aware MBONE applications
  • VAT, NV, etc.
  • Next talk, not this one.

ISPS Router Kernel

• Modified BSD kernel
• Internal packet forwarding path:
  Figure (GIF - 30365 bytes)

Packet Classifier

• Classifies packets into groups for processing
  • Routing
  • Resources management
  • Link sharing
• Derived from Joel Halpern's discontiguous Patricia tree algorithms
• "Single Level" classifier - all decisions are merged into a single classification action for high performance
• Dynamic, in-place (usually...) class merging and unmerging
• Approach has problems when fields can move (variable length, encapsulation, etc.)

Packet Scheduler

• Schedules packets for transmission by an interface.
• Implements the CSZ scheduling algorithm (described in old svc model I-D, later papers)
  • Building blocks for end-to-end services:
    • Guaranteed service
    • Predictive service
    • Multiple priorities of ASAP service
  • Hierarchical link sharing (control of local link usage)
  • Modelled as "intelligent queue"

Scheduler Interfaces

• Queue Interface
  • isps_queue(ifnet *, mbuf *, isps_class_t *);
  • mbuf *isps_dequeue(ifnet *);
• Control Interface
  • Add/delete/modify packet flow
    • used by RSVP
  • Add/delete predictive service class
  • Add/delete hierarchical link sharing class
  • Associate packet selector (filter spec) with flow
    • used by RSVP
• Information Interface
  • Structured method of obtaining stats, counts, delay info, etc. about the queue

Admission Control

• Variant of measurement-based admission control described in Jamin, et.al.
  • Admits or rejects new service requests based on resource availability
  • Decision based on delay and bandwidth constraints
    • Guaranteed flows - purely mathematical bounds
    • Predictive flows - measured behavior of existing flows, stated behavior (flowspec) of new candidate flow
• "Quasi-replaceable" kernel code
  • Well-defined interface, sharply modular implementation
  • Interface parameters may be specific to algorithm
• Approach promising, but unproven

Supporting Tools

• "iclass" - basic configuration tool
  • Create and delete predictive service classes, link share classes, permanent flows, etc.
  • Command line interface
• netstat extensions
  • new netstat commands show
    • configuration state - share classes, etc.
    • dynamic state - packet flows
    • performance parameters for flows
    • admission control parameters

Remote Management Software

• Provides management and configuration control for remote routers
• Accessible GUI:
  • Map-based top-level interface (tkined)
  • Understandable specification of class parameters
• Preserves config state across router crashes
• Implementation
  • User interface and router components
  • SunRPC communication

Implementation

• Implemented in/on BSD Unix
  • FreeBSD (4.4-vintage BSD on PC platform)
    • PPP
    • Ethernet
    • In progress: T1 (Niwot, SDL), ATM (NatSemi)
  • DartNet SunOS 4.1
    • HSI serial interface @T1 speed
    • Ethernet
  • Currently based on pre-pruning multicast code in each case
  • Two versions somewhat different due to internal differences in routing, kernel support code
• Question: what is useful to release?