Performance Implications of Using VPN Technology for Cluster Integration and Grid Computing

For cluster integration and grid computing that bridges organizational network boundaries, potential showstoppers include security policies, firewalls and non-routable IP addresses. Virtual private network (VPN) technology can overcome these hurdles, but what are the performance implications? In this paper, we analyze the performance implications of using the open-source OpenVPN software. Our measurements include NetIO and the NAS parallel benchmarks. The main performance implications are additional latency (0.5 milliseconds) and potentially maxed-out CPU load on the VPN gateways. Effective bandwidth and the execution time of applications that run across VPN-connected clusters was affected only if both (1) the wide-area link is fast (above 83 Mbits/s, given our 1.4 GHz Athlon CPU) and (2) the application is communication-intensive. Moreover, when moving an application from running within a single cluster to running across multiple clusters, we show that besides potential \´\´VPN slowdown" there is "hotspot slowdown": the grid topology can cause traffic hotspots on the inter-cluster links. The more nodes the cross-cluster application runs on, the worse the potential traffic hotspot