| Popis: |
Assessing traversal times is the main concern in the verification of embedded real-time networks. Schedulability analysis, as it provides firm guarantees, is the preferred technique for the designers of critical systems. There are however contexts where it is not economically or technically feasible to develop one, typically when the software and hardware components have not been designed with predictability in mind, e.g. as soon as TCP-based traffic is involved in network communication or when the hardware platform is too complex (e.g. heterogeneous System-on-Chips). In this paper, we study if it is possible to improve the ability of simulation to observe large traversal times, by running many short simulations with appropriately chosen simulation time and varying initial offsets of the stations on the network. The de-facto standard approach to assess maximal traversal times is to run a single long simulation with synchronized node start offsets and to use randomized node clock drifts inside an acceptable range. This approach is known to yield high traversal times but is not parallelizable. We propose an alternative approach consisting in splitting the simulation time over multiple shorter simulations with, optionally, randomized node start offsets. We evaluate the optimization potential of this simple approach on several realistic network configurations by comparing long simulations to aggregated short simulations, with and without synchronized node start offsets. We observe, considering all flows, that this allows a median improvement of up to 21.3% in terms of maximum traversal time observed, for the same simulation time budget. Additional randomization of the node start offsets showed further improvements of up to 4.8% in our experiments. Results from this line of work can be used to estimate the pessimism of schedulability analyses and verify systems for which no analysis is available. |
