RAID stripes data across its constituent disks. In addition, our AIX host permits striping data across RAID arrays. There are essentially three striping models: (i) wide striping; (ii) narrow striping; and (iii) no striping. Wide striping lays out contiguous data across all the arrays whereas narrow striping lays out data across only a subset of the arrays. For a detailed study comparing wide striping to narrow striping, please see [40]. For a very useful practical introduction to the mechanics of implementing striping, please see [41]. For random streams, striping is useful in reducing ``hot spots'' (points of contention) leading to a reduction in the average response time for random seeks.
While wide striping is extremely useful for random clients, it is
not friendly to sequential clients. Because striping is done at
the host level and is not visible to Shark, when wide striped,
each sequential access stream could appear as multiple, although
slower, sequential access streams. Narrow striping is
likely to have moderate number of hot spots for random clients,
but is friendlier to sequential clients. In this paper, we have used narrow striping that stripes across
RAID arrays. Our insights, analysis, and algorithm do not
change with wide striping or with no striping.