Linux Scalability -- From the Micro to the HUGE
Linux is possibly the most scalable operatings system *ever*. From the Linux Watch (http://researchweb.watson.ibm.com/WearableComputing/factsheet.html) to the 1024 and more processor ALTIX machines from SGI; one kernel runs on them all.
The kinds of challenges that are presented by the very small are in some way related to the challenges of the very large, but for different reasons. For example, memory footprint is very important for both: on small systems you just don't *have* much memory; on large systems, caching effects dominate performance --- if you can keep cachelines local, you win.
Likewise, power management is extremely important for both ends of the scale. At the low end, to maximise battery life, at the high end, to minimise airconditioning costs.
However, the mechanisms available, and the specific goals differ.
In this talk, I shall attempt to explain what scalability *is* as far as an operating system kernel is concerned, explain where Linux has come from and is now, and then give a possible vision for the future, as to how Linuix can maintain and enhance its already good scalability. I'll also present benchmarks on a variety of machines to show where the problems are now.
After getting his PhD in distributed operating systems from UNSW, Peter hacked various Unix kernels for 13 years at Softway, working amongst other things on checkpoint-restart and fair-share resource management.
He now works with Gelato@UNSW, which is a key member of the Gelato federation scalability focus group. Gelato is about performance and scalability for Linux on Itanium.