A Hybrid Testbed for Performance Evaluation of Large-Scale Datacenter Networks

Abstract

Datacenters (DC) as well as their network interconnects are growing in scale and complexity. They are constantly being challenged in terms of energy and resource utilization efficiency, scalability, availability, reliability and performance requirements. Therefore, these resource-intensive environments must be properly tested and analyzed in order to make timely upgrades and transformations. However, a limited number of academic institutions and Research and Development companies have access to production scale DC Network (DCN) testing facilities, and resource-limited studies can produce misleading or inaccurate results. To address this problem, we introduce an alternative solution, which forms a solid base for a more realistic and comprehensive performance evaluation of different aspects of DCNs. It is based on the System-in-the-loop (SITL) concept, where real commercial DCN equipment (switches) is interconnected with simulated DCN nodes via the SITL virtual gateway modules in the Riverbed Modeler, forming a unified hybrid real-simulated DCN setup. This testbed is complemented with high performance network testers, which can generate a mix of different real traffic streams at L2-3 and L4-7 at the rates of 1, 10 or up to 40 Gbps. These components allow performance benchmarking tests to be conducted at large scale.

Analysis of traffic engineering capabilities for SDN-based data center networks

“…In addition, we also present a hybrid (physical-simulated) testbed (for Proof-of-Concept please refer to [30]; a system without SDN capabilities is presented in [31]) for evaluating SDN TE characteristics at scale. We are using the hypercube structure illustrated in Fig.…”

“…Fig. 5 A simulation model using an Incomplete Hypercube structure with 44 nodes and links to external DCN equipment [31] V. CONCLUSIONS In this work we have discussed the key limitations of the traditionally used TE approaches, taking into consideration the operational performance requirements in DCNs, due to completely different traffic profiles, with burstiness, critical bandwidth or ultra-low delay requirements, in addition to the sub-second lifetime of the dominant volume of all the traffic flows. These limitations include non-optimal path computation algorithms, slow TE Database convergence times, which is a critical requirement for the real-time control plane decisions.…”