G-WAN is faster on Linux (Ubuntu) than on Windows (Vista)

HTTP Keep-Alives:

No HTTP Keep-Alives:

Why G-WAN/Linux does not dwarf all other HTTP servers like on Windows?

• Linux does not slow-down applications like Windows does with its 300% user-mode overhead; this explains why everybody is faster on Linux, and why, on Windows, all (but G-WAN) are slower than IIS 7.0 in the kernel;
  
  
• Linux is faster, but it has also its limits. As G-WAN takes much less CPU time than the kernel to operate (all other servers consume much more CPU time), the Linux kernel is a bottleneck only for G-WAN.

This explains why the gap between G-WAN/Linux and other servers is smaller than on Windows (where reaching kernel's performances is much harder).

At concurrency=1000, G-WAN uses only 4% of the CPU** while Rock needs +90% of the CPU (22.5x more) to process 1.6x less requests.

Also, getrusage() tells that G-WAN uses only ~1%** of this 4% CPU load while the kernel uses the other 99%. If G-WAN was 10x slower, its performances would be identical (it would not be the case for Rock).

With a 90% CPU load, Rock takes (90-4)/4=21x more time than the kernel and is therefore thousand times (21x100) slower than G-WAN. Linux only shows a 1.6x difference between G-WAN and Rock because, while Rock is slowing-down the kernel, G-WAN is slowed-down by the kernel.

Would the kernel let G-WAN saturate the CPU, G-WAN would theorically process 25x (4%x25=100%) more requests per second (while Rock could only add 10% to its score since it already uses +90% of the CPU resources).

Well-known Hardware and OS designs erected this wall. G-WAN just reveals how large the CPU/memory and CPU/IO progression margins are.

Solaris will tell us (soon) how much an OS can make the difference -and if a better designed hardware platform is required to let G-WAN saturate a CPU.