The Kalray MPPA2 manycore processor implements a clustered architecture, where clusters of cores share a local memory, and communicate through a RDMA-capable network-on-chip (NoC). This NoC has been designed to allow guaranteed delays by the adequate configuration of traffic limiters at ingress and the choice of routing. We first present the challenges related to the routing of concurrent flows and a strategy that solves it. Routing challenges include deadlock-free routing, which is always an issue on wormhole switching networks, fairness of resource allocation between flows, and ensuring the feed-forward property required by deterministic network calculus (DNC). Second, we present a linear formulation based on DNC for computing end-to-end delay bounds of concurrent feed-forward flows on the MPPA2 NoC. Finally, we compare this linear formulation to a classic formulation originally designed for the AFDX avionics networks.