Figure 2

Scaling relations for Hadoop I and II with increasing input dataset size, from S1 to S9. The figure highlights the almost linear scaling characteristics of the Hadoop approach. Our MR implementation of the preprocessing stage (A) displays almost perfect linear scaling. The ‘ideal map’ curve for Hadoop II cluster (open circles) shows (B) the deviations from the real case (closed triangles) due to map jobs failures; that is, additional time to reallocate and restart jobs. The SNP calling stage (C) scales worst, compared with the others, mainly because of the poor job balancing of the Crossbow.