Network Function Virtualization not only can help service providers to offer network services flexibly, but also can reduce cost of supplying space and energy for physical devices. User requests may demand different numbers of services in a chain, in which the related flows can be directed to different machines hosting required virtualized network functions (VNFs) in a designated order. As a result, service providers need to deploy services with limited resources to fulfill requirements from various users. Conventionally, these requirements are represented as service level agreements (SLAs) in terms of CPU, memory, bandwidth, latency and etc. However, current service deployment solutions often target operation/energy cost, neglect network cost (e.g., network congestion and overdue latency) and/or cannot run in real time. Therefore, we propose a new approach called SLA-driven Ordered Variable-width Windowing (SOVWin). In SOVWin, resource capacities of requests/services are sorted to accommodate user requests as many as possible and a windowing technique derives a legitimate service deployment that minimizes transmission latency as much as possible. Our experiment results indicate that not only can SOVWin run efficiently and fulfill SLAs successfully with a high user acceptance rate (99.9%) under different number of services, but also can reduce 95.7% latency violations induced by a baseline (first-fit) approach.