On the near horizon are SSD products that cost between $10 and $20 per gig and
have read latencies in the neighborhood of half a milisecond. It would be nice
to be able to leverage this type of storage in situations where the working set
was bigger than could fit into RAM (at least economically). One possibilitiy is
to have a "second tier" of ARC that would use one or more SSD. This could provide
a huge performance win for workloads with large working sets or huge cost savings
for workloads with more modest working sets and more relaxed latency requirements.
This is PSARC/2007/618