The cloud-native paradigm advocates agile development and deployment of virtualized micro-services, introducing a flexibility and dynamicity for service endpoints that may exist in many locations of a provider's network, not just data centers.Such ability leaves open the problem of scheduling traffic from clients to those possible locations. In this paper, we position our solution to this problem at the data plane level, avoiding the shortfalls of existing solutions in terms of latency and path stretch. For this, we present a system model for forwarding service requests based on compute information, with a distributed scheduler realizing the traffic steering decision at line rate and with measurable performance gains against existing networklevel solutions. We evaluate our solution against several design aspects to provide insights for real-world deployments, while quantifying performance improvements for use cases where such scheduling decisions could indeed be performed at the level of each service request. Here we show that our improvements in request completion time may lead to serving up to 162% more clients within the bounded request time that would ensure acceptable quality of experience.

Service-level traffic steering in the Internet has been using an indirection-based model for decades now, using the DNS to resolve a name to a locator, often complemented with load balancing techniques. Contrasting this off-path realization, service information as part of the data packet itself may determine the one of possibly many communication endpoints on-path while traversing the network. This paper compares both design choices regardless of the specific decision mechanism used. For this, we assume a compute-aware traffic steering mechanism for both approaches and determine latency penalties through off-path resolution steps as well as distributing scheduling decisions to on-path network ingress points. Lastly, we investigate latency variances and resilience in an AR/VR scenario.