Lists Home |
Date Index |
> Considering that XSLT/XQuery is orders of magnitude more complex than
> simple crypto algorithms, I'm wondering to what extent hardware
> addresses compliance, implementation bugs, upgrades, dynamic callback
> features, effective programming language integration, etc.
I've always been extremely skeptical about custom hardware for this kind of
application, firstly because the effort of designing it detracts from the
effort available for producing better algorithms, and secondly because once
you've built it, it's much harder to change than a software solution.
Early in my career we did a study looking at implementing some core parts of
a database management system in the mainframe microcode. We ended up
concluding that for a ten-person-year investment, we could achieve a 20%
speed-up. But we could achieve a 40% improvement by spending 3 person-years
on improving the hardware and 1 person-year on improving the software,
without the disadvantage of coupling the two together.
Any performance improvement that is achieved by interfering with the
architectural layering of the system is likely to be short-lived, because it
restricts your ability to evolve each of the layers independently. This
doesn't only apply to doing things in hardware.