Data Layout

Traditional disk data layout minimizes seek time and rotational latency. Analogously, we chose a layout for probe-based storage that has a similar sequential ordering to minimize movement for consecutive requests accessing sequential data.

Figure 2: Example data layout over two movers in one MEMS device.

Figure 3: Data access for a single mover with four clusters and one tip per cluster. (a) Tip 1 is activated and the mover slides up, reading the column of bits the tip passes over. (b) Tip 2 is activated after the shaded bits have been read, and the mover reverses direction. (c) After all columns have been read, with an $x$-move between each column to reposition the tips over the new column of bits, tip 3 is activated and the mover slides up, analogously to (a). (d) Tip 4 is activated and the mover reverses direction, analogously to (b). (e) All bits in the mover are read.

(a)	(b)	(c)	(d)	(e)

Each mover is divided into several clusters, each of which is an area covered by a grid of tips. For each cluster, only one read/write tip can be active at a time. This cluster design helps to minimize power consumption while increasing the size of the swept area. Given this parallel architecture, where several tips can read/write data in parallel, it was reasonable to stripe the data between the clusters. We chose to work with bit-level striping because it maximizes the throughput when there is only one outstanding request.

Figure 2 shows the layout of a MEMS device with two movers, two clusters per mover and four tips per cluster. This makes the tip area a quarter of the cluster's area. One tip per cluster may be active at a time; therefore, in this device, four tips are active at a time. Consequently, a sector of data is read/written by the four active tips simultaneously. The sector is bit-interleaved between the two clusters, as depicted by the pattern of the bitstream in Figure 2. The data is accessed while the mover is moving in the $\pm y$ direction. Figure 3 shows how data is accessed on a single cluster.

A single device may contain several movers, increasing parallelism. To exploit this secondary level of parallelism, we stripe data bitwise among all the clusters of these movers. For example in Figure 2 there are four simultaneously active tips, one per cluster, and the bitstream is striped among them all. First the stream is split into groups of four bits each, as depicted by the different gray shades. Then each such group of four bits is interleaved between the four active tips, as shown by the arrows in Figure 2.