Derby SMP Scalability - our report

[Anders Morken <an...@stud.ntnu.no>]

As we've mentioned on the list earlier, we (I and Per Ottar Ribe Pahr)
spent our last semester looking at Derby on multiprocessors and writing
a report on what we found. We also promised to post it here in the hope
that it is useful to someone, so here it is:

http://folk.ntnu.no/andersmo/derby_project_report.pdf

I believe the general audience on derby-dev can skip chapter 2 without
missing out on anything they don't already know. =)

There are some things that didn't make it into the conclusion - most
importantly this:

B-tree latching is a significant bottleneck when the number of CPUs
available grows to 8, and we believe it significant on 2 and 4 CPUs as
well. Section 5.3.1 discusses this. The root block of the B-tree stands
out as a significant contention point, see table 5.2. (The source data
for said table actually included both latches and all regular locks in
the lock manager - but nothing but latches made it into the top ten. =)

We're currently looking more into at this very specific problem. Maybe
we'll pass some more findings and suggestions on to you in a new report
this summer - and any preliminary findings we believe to be significant
might also be posted here on derby-dev.

Thanks,
-- 
Anders Morken

My opinions may have changed, but not the fact that I am right!