Even more important than overlapping expensive computation (on SMP
machines only) is covering I/O latency on all machines.  This will be
especially important on servers with many spindles, lots of RAM, fast
networks, and lots of clients.

I bet a lot of the benefit could be realized by putting SFTP I/O on
separate kernel threads and reviving RVM's multithread support to allow
the log truncate to run concurrently with updates.

As for portability, from some perspectives using Pthreads is a big win,
but I'll say that getting LWP running on W95 took only a couple days
precisely because LWP is so simple and relies on nothing but a working
select().

Michael