The amount of BS could range from none to complete; it all depends upon your access pattern. (my workload has heavy recent locality, so I could probably even l

The amount of BS could range from none to complete; it all depends upon your access pattern. (my workload has heavy recent locality, so I could probably even live with Jim Gray's predicted combination of main memory + tape; where the SSD buffer would come in handy during reboots) What does JNG's five-minute rule suggest for lifetimes and xfer sizes in 2024? https://www.cs.cmu.edu/~natassa/courses/15-721/papers/gray.pdf https://www.cs.cmu.edu/~natassa/courses/15-721/papers/gray.p... (1997) EDIT: with an SSD at 40k iops @ 100 USD and 4k pages: (256 pages/MB) / (40000 io/s) * (100 USD) / (1e-6 USD/MB) ~= 1 week / io Randomly Accessed Objects => 1 week rule assuming sequential b/W at 6 GB/sec: (1 / 6000 MB/sec) * (100 USD) / (1e-6 USD/MB) ~= 5 hours Sequentially Accessed Objects => 5 hour rule I'll let someone else do the index calcs, but clearly we're no longer in a world where secondary swap makes any sense during typical user process lifetimes.