Course notes ELG 7187C - Modeling hard real time and performance

Different types of requirements and tools

• Different non-functional requirements
  • Response time: time delay between input and corresponding output interaction
    • Assumption: throughput (number of requests per second), interarrival time delay
      • Note: response time depends on throughput - there is a "maximum throughput"
  • Reliability: operational period until the system fails
  • Availability: Probability that the system is up and running (depends on reliability and time period required for the repair)
• Different nature of requirements
  • hard real-time constraints: there are maximum and minimum bounds for delays, such as response time, inter-arrival time, operational period, etc.
  • stochastic systems: we obtain statistical properties, delays are characterized by distributions (over time delay)
    • different assumptions about inter-arrival times - typical assumptions: fixed, Poisson (exponential), long-range dependencies (burstiness)
    • typical kinds of distributions (to simplify the formulas and calculations)
      • fixed delay (delta-distribution)
      • exponential (Markov processes)
      • Normal distriburtion (for simplifying folding - addition of delays)
      • uniform distribution between min and max
      • arbitrary distributions (Semi-Markov processes)
• Different modeling tools
  • for hard real-time modeling:
    • Timed Automata : see tutorial slides and demo of Bridge example using the UPPAAL tool (see also some documents C )
  • for statistical reliability modeling:
    • Markov models (exponential distributions for transition firing times): Note: there are simple analytical methods for obtaining performance parameters, however, the number of states may be big for realistic models. See introduction A.
    • Semi-Markov models (general distributions for transition firing times).
    • Generalized Semi-Markov models (several independent timers, running to zero, that can be set to an initial non-zero starting time which has that general distribution; when they reach zero, they trigger a transition). See example C (from the PhD thesis by Fda Dankar)
  • for statistical response time modeling - taking shared resources into account
    • Queuing systems : The waiting time in the request queue of a server is determined as a function of the service time of the server and the inter-arrival time distribution of requests arriving at the server. Queuing networks involve several servers with their respective queues and a workflow pattern that indicates the flow of requests through these servers. See introduction.

Modeling approaches

• Global time variable (e.g. variable NOW in SDL)
  1. to observe response time: keep copy of Now in variable x when request arrives, check (x - Now) when response is provided
  2. enabling condition depending on time for a transition (e.g. abandon action if it is too late)
  3. condition with urgency: forces transition to occur before the urgent condition becomes false (if such a transition is possible)
  4. timing conditions in UML sequence diagrams (real-time profile): introduces (a) time variables corresponding to the sending or reception of messages, and (b) conditions about relationships between the values of these variables (like under points 1 and 2 above)
• Transitions take time. This approach can be added to LTS, Petri nets, IOA, FSMs
  • when entering a state the decision of the next transition is immediate
  • for performance measures, one has to define probabilities for the different transitions from each given state
  • transition times may be fixed or varying (delay distributions)
• Selecting the next transition from a given state takes time. In this case, the execution of a transition is usually assumed to be immediate. Also, the probabilities o the different transitions from a given state must be given. This approach can be used with LTS, Petri nets, IOA or FSM. This allows the modeling of hard real-time systems (e.g. Timed automata) and stochastic systems (e.g. Markov models) - see above.

Standardization: There is a UML profile for Modeling and Analysis of Real Time and Embedded systems (MARTE) that deals with these issues. << Some performance analysis tools have been developed in conjunction with UCMs based on a Common Scenario Model (common to UCM and Activity Diagrams) for performance modeling (see Petriu, D.B., Amyot, D., and Woodside, M. (2003) Scenario-Based Performance Engineering with UCMNav. 11th SDL Forum (SDL'03), Stuttgart, Germany, July 2003. LNCS 2708, 18-35) >>

Last updated: February 22, 2013