Figure 2: Speedup of four- and eight-thread checkpoints relative to user-level memory copy.