3:1 4:4 100HZ 1000HZ comparison with the HINT benchmark

This compares with the HINT benchmark 4 kernels sharing the same 2.6.5-aa4 codebase but compiled with different settings, the raw numbers generated by the benchmarks can be downloaded from the below links:

3:1 user/kernel split with HZ=1000 (same as 2.6 vanilla) - DOUBLE-2.6.5-aa4-1000
3:1 user/kernel split with HZ=100 (same as 2.4 vanilla) - DOUBLE-2.6.5-aa4-100
4:4 user/kernel split with HZ=100 - DOUBLE-2.6.5-aa4-4:4-100
4:4 user/kernel split with HZ=1000 - DOUBLE-2.6.5-aa4-4:4-1000

The 4:4 kernel patch is merged from 2.6.5-rc3-mm4.

In very short the HINT benchmark simulates some realistic number crunching purerly userspace workload at different working set sizes.

I'm plotting the working set size on the X axis, and the memory bandwidth generated by the CPU using each working set size on the Y axis. The higher the memory bandwidth at any given working set the better.

Some comments are embedded in the images.

Conclusion

Depending on the number crunching working set the 4:4 option can be a double digit performance hit even for pure userspace load. (this happens at exactly 500kbyte working set for my hardware)

Depending on the number crunching working set HZ=1000 may generate up to a -7% slowdown compared to HZ=100 if the 4:4 option is enabled. (this again happens exactly at 500kbtye working set for my hardware)

If the memory footprint is below and above a certain threshold (below 100k and above 1M for my hardware) 4:4 or 3:1 doesn't make any difference for purerly userspace load with not more than 1000 irqs per second and zero syscalls and zero page faults.

Obviously with different tlb/cache sizes and cpu variations the threshold numbers will vary.

These results are 100% reproducible for me. The HINT is under the GPL so you should find it on the internet (if not feel free to ask me for a copy).

andrea@suse.de