In the late 1960s, land surface schemes were devised to represent heterogeneity of land surface as the lower boundary condition of atmospheric models. They were originally designed to reproduce realistic energy partitioning and water balances within a simple structure, but many functions have been added to consider various underlying processes such as the carbon cycle. Meanwhile, there also have been remarkable developments in the field of observations. Advancements in methodology, instrument sensor technology, and observational networks have improved precision and spatiotemporal coverage, and, especially, satellite remote sensing has provided continuous global observations since the late 1970s. Despite those advancements in models and observations, large-scale land surface simulations are still struggling with uncertainties originating from various sources including human-induced artifacts in numerical implementations and the uneven distribution of observation networks. Further, all models and observations have their own strengths and weaknesses, subject to their key objectives. Therefore, a test-bed is needed to diagnose the current status of state-of-the-art models and input observations, and to evaluate simulations of various spatiotemporal scales in a component-wise comparison using a broad set of independent observations. The second phase of the Global Soil Wetness Project (GSWP2) produced the first multi-model ensemble analysis of global land surface state with 13 land schemes for 10 years period (1986-1995). In the third phase of GSWP (GSWP3), it is proposed to investigate the long-term changes of the interactions between the components of energy-water-carbon cycles with extensive data and model verifications in ensemble land simulations. Here, we present some lessons learned through the preliminary experiments of the GSWP3.