LRVM (Satyanarayanan, et al.)

Good points:

1) Providing an abstraction of a greatly needed behavior
(transactions) makes system code implementation much easier: this stuff is
useful.

2) Returns to UNIX mentality of small and simple building blocks.

3) Performance analysis (Rmem/Pmem) very applicable to stated domain (fs
metadata).


Bad points:

1) It would have been nice if they had explicitly stated that set-range
can be called multiple times within a transaction; they only comment on it
in 5.2 when discussing optimizations (for overlapping region
specification).

2) It's unclear why the throughputs are almost equivalent for sequential
access even though their CPU utilization is much different. This seems to
contradict their scalability concern, as it would seem both systems are IO
bound as opposed to to CPU bound; given the rate of CPU improvement, IO
would seem to be the greater concern. Of course, it's still good that the
very simple RVM performs better.