Hi, I'll apply the patches and rerun our tests today. Thanks guys! -gary
On Sun, Jan 24, 2016 at 5:49 PM, Willy Tarreau <[email protected]> wrote: > Hi again, > > On Sun, Jan 24, 2016 at 09:00:31PM +0100, Willy Tarreau wrote: > > OK so I'm having a preliminary patch for the performance regression > > causing incomplete buffers to be processed. It's not optimal yet, > > but it's safe and works for me. Now I can reproduce the tests and > > always observe the exact same performance between all runs +/- 0.2%. > > I've cleaned up the patch and implemented the enforcement of the exact > requested buffer size. Now the default setting being 16384, it leaves > in a single SSL_write() and the performance issue is gone. I'm getting > 416 req/s stable. I'll backport these patches to 1.6 but first I want > to understand why we're seeing a freeze at exactly 16000 bytes. We have > it in 1.5 too. > > I'm attaching the 3 patches to apply for Gary. They should apply > cleanly over 1.6. That's all for this evening or I'll start doing > stupid things. > > Willy > >
