Performance analysis of master-slave multi-hop switched ethernet networks

Abstract

There is an increasing trend towards using switched Ethernet in real-time distributed systems due to features like absence of collisions and high throughput. Nevertheless, a few problems persist, in particular related to priority inversion and limited length in queues. In this paper we focus on a protocol which uses a master-slave technique over standard switched Ethernet in order to overcome such problems, namely the FTT-SE protocol. We present an improved response time analysis for such a network and we compare, analytically and with simulations, the results achieved with Network Calculus on a worst-case scenario. We show that our proposed response time analysis gives tighter bounds compared to Network Calculus. Moreover, we compare the performance of different solutions to scale the FTT-SE protocol with respect to the bandwidth utilization. Finally, we propose a new architecture to improve the average performance of master-slave switched Ethernet networks.

Allocation of Parallel Real-Time Tasks in Distributed Multi-core Architectures Supported by an FTT-SE Network

“…In this paper, we assume the network analysis presented in [21] for FTT-SE networks. Thus, for a partition of tasks τ i to be considered a valid solution (all deadlines are met), the following condition has to be respected:…”

“…Depending on a given partition, we have to find the WCRT of the messages in the network to verify if the condition in (13) is respected. We consider the work presented in [21] for the computation of the WCRT of messages within the FTT-SE protocol, with a slight modification. The request bound function rbf i,j,k (t) represents the maximum transmission requirements generated by a message µ i,j,k and all its higher priority mes-sages during an interval [0, t].…”

“…For computing the switching-delay-effect Is i,j,k (t), it is needed to compute an upper bound on the number of switching delays (SD i,j,k ) from each message that contributes to (15), at time t. In [21], depending on time t, a number of switching delays are inserted into an array whenever a message crosses a switch in the network. The array is sorted in order to consider the maximum switching delays only.…”

“…A message µ i,j,k can be blocked by other higher priority messages even if they do not share a transmission link. Thus, a higher priority message can delay a lower priority message even though they do not share a transmission link [21]. Therefore, to compute the worst-case remote link delay, it is needed to consider all messages that share links with the messages that contributed to the shared link delay (see (15)), excluding all messages that are already considered in (15).…”

“…If threads θ i,j,k are executed remotely, the WCRT of messages transmitted through an FTT-SE network has to be considered. That is modeled with (20)- (21). Finally, all tasks have to respect the condition in (13).…”