OSNews

Hi,

For HPC there's a tendency to split a large amount of work into "one piece per CPU". This makes the scheduler mostly irrelevant (it's simple to decide which piece of work the CPU should be doing when there's only 1 piece of work the CPU can be doing). The performance difference probably comes from something else (e.g. maybe Linux has more efficient networking).

When you're running a wide variety of different tasks (with different characteristics), then the scheduler becomes important. I've said for a long time that the scheduling in Linux is flawed, because most processes can't tell the scheduler anything about their characteristics, and therefore the scheduler doesn't have the information needed to make good decisions.

In contrast, there's lots of OSs that are better for desktop interactivity than Linux (including Windows) because tasks/processes do tell the scheduler something about their characteristics. For example, for Windows there's a "priority class" and a "base priority" within that class. For Linux there is a "scheduling class" (but the same scheduling class is used by almost everything and doesn't allow for the equivalent of a "base priority").

This patch helps to hide a symptom of the problem (and helps to illustrate how bad the problem is), but hiding symptoms isn't the same as finding a cure.

- Brendan

I think I know why. It's because on a machine that probably needs millions of dollars just to flip the switch, there are more important things than just nailing a benchmark. I don't know if Linux is more efficient than Windows in HPC, but the article you linked doesn't say much either. "

http://www.top500.org/stats/list/36/osfam

Linux = 91.80 % share of supercomputers, which are the world's most expensive of machines, where only performance matters.

Windows HPC = 1%.

Draw your own conclusions.