Network Function Virtualization (NFV) is a new trend in networking, where network functions are moving from custom hardware appliances to software implementations running on virtual machines (VMs) hosted on commodity hardware. While the benefits of NFV such as cost reduction and increased agility are well understood, doubts still exist on whether a software implementation can match up to the high performance that hardware appliances deliver. In this context, network operators would benefit from frameworks that monitor performance and identify bottlenecks in Virtual Network Function (VNF) implementations obtained from vendors. While several techniques already exist to identify performance issues in cloud-based applications, most of them either use hardware resource utilizations to identify hot-spots (making them incapable of detecting non-hardware performance bottlenecks) or rely on application specific measurements (which may not be exposed by VNFs to vendors always). This paper describes NFVPerf, a performance monitoring and bottleneck detection tool for NFV. NFVPerf works as part of a cloud that hosts a NFV deployment, and takes a configuration file specifying the basic architecture of the VNF as input. It sniffs packets on all VM-to-VM communication paths, computes per-hop throughputs and delays, and uses these “black-box” measurements alone to identify performance bottlenecks (including software bottlenecks) in real time, without requiring any instrumentation of the VNF. Further, NFVPerf can be customized to any VNF implementations with just configuration changes. Our evaluation of NFVPerf shows that it can monitor performance and identify bottlenecks in an NFV deployment, with high accuracy and minimal overhead. We believe that a system like NFVPerf would form a great addition to cloud management systems in the era of NFV.