Many of the algorithms that we use today were designed to improve performance on machines that were much smaller and slower than today's machines. The analysis of, and justification for these algorithms usually makes assumptions that are no longer true for today's machines processing large datasets. Examples of these assumptions are that all memory references are equally costly, that memory access doesn't matter in analyzing algorithms, and that caches make small strides more efficient than large strides. The talk provides more questions than answers, but shows some empirical data that raises a lot of questions about traditional data structure design for large data.