Evaluating the Performance of Sidecar-based and Sidecarless Cloud-native Service Mesh Solutions

The software development landscape is rapidly evolving, marked by the shift from monolithic applications to microservices architectures and the adoption of agile methodologies. While developing microservices has become more effortless, the challenges of maintaining and deploying them persist, specifically with managing the complex dependencies associated with multiple microservice deployments. With an increasing demand for container orchestrators and Kubernetes emerging as the industry standard, networking microservices remains a key challenge. Service mesh, a solution to this problem, consists of a control and a data plane, with the former responsible for managing the components of the latter. Data planes in most service meshes are implemented using sidecar proxies, posing specific challenges when injected into application pods. A potential solution to these challenges lies in the sidecarless approach. This paper aims to explore the eBPF-based service mesh, Cilium, which eliminates the need for a sidecar while encompassing most service mesh capabilities. Further, we benchmark Cilium’s performance, analyze its effectiveness, and compare its performance with Istio and Linkerd. Our evaluations show that Linkerd outperforms other solutions with respect to response times, with Linkerd performing 29.85 % and 63.43% better than Cilium and Istio, respectively.