TLM - Can it estimate performance and utlisation ?

Q. If we do not model clock cycles or instruction counts, can we work out response times and resource utilisations ?

A. Attach useage annotations to whatever behaviours are actually modelled and add them up!

Can use fluid-flow continuous approximations to bus-bandwidth, cache-capacity, DRAM response time and so on.

But most important thing to keep track of is elapsed time.

Modelled behaviours:

• Memory transactions by processor: annotate with average (minimum) cost.
• Instructions by processor: annotate with time used on an average instruction (or per instruction) basis.
• Packets or flits in network on chip.
• Contention overheads: can use geometric (1/(1-p)) approximations.
• Clock-cycles: where cycle-accurate sub-systems are included.

We have big choices in the way contention and transaction ordering are modelled:

• accurate, non-blocking model: fully simulated,
• approximate blocking/non-blocking: approximately simulated,
• temporally-decoupled, loose timing: largely ignored.