If you're crashing you're probably running out of swap space, so
you'll want to look closely at the memory consumption of each site's
nsd process. Also take a look at your postmaster -B switch value. On
my 256MB dual P450 (faster than Ben's, nah-nah-nah!) I've got -B set
to 8000 16KB blocks, i.e. 128MB or half the machine. I'm only running
two sites, though, which hover around 90 MB in consumption combined.
The two sites aren't very busy, but one's very database intensive,
thus the large buffer cache for PG.
I'll be interested in seeing Janine's report on memory consumption
after turning off fastpath on one of the sites that has a lot of
static pages and a fair amount of traffic. The idea is that turning
off fastpath will be slower, but not too bad because most of the
resulting filesystem reads will come from the linux file buffer cache.
Fastpath takes a lot of RAM because each thread has its own cache,
while the linux file buffer cache is shared systemwide.
For a relatively unbusy site with lots of static pages you should be
able to decrease RAM consumption this way. For a busy site, of course
you'd want to pack the machine with as much RAM as possible and turn
fastpath loose with a fairly high per-thread cache limit. Of course,
you probably wouldn't be running 10 sites on the same machine as the
monster site, either! Different approaches for different scenarios.
Of course, Janine may come back in a day or two and report that this
approach didn't help at all. We'll see!
