A system in which convection takes place in the upper mantle on two distinct horizontal length scales is proposed. This is consistent with the existence of the plates themselves, the relatively constant heat flux background in older ocean basins, and the knowledge of convection in fluid layers gained from laboratory and numerical experiments. The large-scale circulation consists of the plates themselves and the return flow necessary to conserve mass. The small-scale flow, analogous to Rayleigh-Benard convection or variants of this, which have been the main target of numerical study, provides the necessary vertical heat transport in the upper mantle that supplies the required heat flux at the base of the lithosphere. The depth of convection is taken to be down to the 650-km seismic discontinuity, and this depth characterizes the horizontal length scale of the small-scale convection. This system is studied by means of a set of laboratory experiments that explore the interaction of the small-scale convection with the large-scale flow. The experiments show the plausibility of convection on two scales. Furthermore, they suggest that beneath fast-moving plates (absolute velocities around 10 cm yr-1) the small-scale convection will align itself as rolls in the direction of the large-scale flow in geologically short times. However, beneath very slow moving plates the times required for the alignment of convective rolls are long in comparison with times over which no changes in plate motions are to be expected. Here the convective platfrom is more likely to take the form of upwelling and downwelling spouts. Thus this simple system of a convecting layer beneath a moving boundary contains the possibility of explaining a wide variety of surface features. Observational tests of the consequences of the two-scale idea are suggested, and the assumptions on which this idea is based are critically discussed.