Two-dimensional PIV was used to measure the cycle to cycle variability of large-scale flow-structures at TDC in a motored, two-valve, four-stroke engine. Over two hundred velocity distributions were measured for both a highly directed flow using a shrouded valve and a relatively undirected flow using a standard valve.Each cycle of the directed flow had the appearance of the single large swirl structure seen in the ensemble mean. Cyclic variability of the large-scale flow structure was manifest as variations in swirl ratio (rotational speed). Generally the variability was limited to scales smaller than 10 mm in size.For the undirected flow, none of cycles had the appearance of the ensemble mean. The flow appeared to be multimodal in that large-scale flow-structure patterns could be classified into three types based on flow-pattern recognition. This multimode behavior of the cyclic variability resulted in velocity probability density functions, PDFs, that were nongaussian in shape and in some regions of the cylinder were themselves multimode in appearance. It is demonstrated that small cycle-sample size can result in extreme statistical bias in the ensemble mean.