> For an operating system latency is more important than through put
One interesting exception to this rule is operating systems for network packet routers. These need to be optimized for throughput over latency. I mean, the difference between 10 and 20 million packets per second is important, but the difference between 10 and 20 microseconds of processing delay is not.
This is a fun area to be in at the moment because simple academic-looking code can actually compete with highly optimized kernels like Linux because those have been optimized for something else.
When you're talking about an average latency of single digit us for a regular switch and < 500ns for a low-end performance switch, the difference between 10 and 20 microseconds is massive.
Packets between internet hosts spend about 10,000,000ns in transit and they traverse around $10,000,000 worth of network elements on the path. These numbers are both worth optimizing. I'd say that the lowest hanging fruit for operating system hackers like myself is the cost i.e. serving more customers with fewer boxes. That latency figure is tougher because it's mostly due to the speed of light rather than processing time.
One interesting exception to this rule is operating systems for network packet routers. These need to be optimized for throughput over latency. I mean, the difference between 10 and 20 million packets per second is important, but the difference between 10 and 20 microseconds of processing delay is not.
This is a fun area to be in at the moment because simple academic-looking code can actually compete with highly optimized kernels like Linux because those have been optimized for something else.