openssl performance delta built-in vs custom compiled

[Marcus Pollice]

On 21.12.2014 06:26, Dimitri John Ledkov wrote:
> Note that these days compilers optimise C at times much better than
> hand-crafted (crippled) assembly.

I'm aware of this, but it's still not universally true. Also this would 
have to go to the upstream developers.

> Plus if one has compatible CPUs the hardware engines kick in thus
> making comparison harder (e.g. AESNI, RDRAND etc.)

Correct but not applicable, since I compared the performance on the same 
machine using the same version, with these features enabled in both 
cases. Also since I focused on hash algorithm performance it wouldn't 
have changed something.

> You are not showing your build log, the chroot used to build and the
> package version of a toolchain used.... Note that launchpad.net
> publishes full build logs for all binaries including the host kernel
> and toolchain packages used. Note that binaries are preserved in
> debian/ubuntu and not rebuilt from scratch with each kernel / gcc /
> binutils uploads, thus one can have binaries published in a stable
> release built with a toolchain that's no longer available in the
> archive. Thus you need to compare the toolchain you are using with the
> one used to build the corresponding openssl binary package you are
> trying to match on a same architecture.
>
Thanks for pointing me into this direction. From a first glance on the 
buildlog I can infer that the same gcc release was used for the package 
in 14.04 as I used for compiling. I didn't use a chroot at all, I just 
downloaded the tarball, extracted, configured and compiled. It's fairly 
easy to replicate this process using a live image. I also didn't use the 
Debian/Ubuntu patchset, which is what I will look at more closely now. I 
suspect that maybe one of these patches changes something that affects 
performance for small data sizes. Is there some documentation available 
somewhere which patch changes what specifically?