2016-09-22 18:29 GMT+03:00 Rémy Maucherat <r...@apache.org>: > > 2016-09-22 17:13 GMT+02:00 Violeta Georgieva <miles...@gmail.com>: > > > Hi, > > > > 2016-09-14 20:43 GMT+03:00 Rémy Maucherat <r...@apache.org>: > > > > > > 2016-09-13 20:07 GMT+02:00 Violeta Georgieva <miles...@gmail.com>: > > > > > > > I would like to back port these changes to Tomcat 8.5 if you do not > > have > > > > any concerns. > > > > > > > > -0.1. I think this is ok, but not certain yet. > > > - Performance is ok for the "common" unencrypted scenario, but it could > > go > > > down if using SSL, for example [= if the need for direct buffers could be > > > hurting - if it is, then there's a problem since SSL will be used almost > > > all the time moving forwards]. > > > > I executed many combinations of performance scenarios - using direct > > ByteBuffer/non direct ByteBuffer, > > HTTP/HTTPS, sslImplementationName - > > org.apache.tomcat.util.net.jsse.JSSEImplementation/org. > > apache.tomcat.util.net.openssl.OpenSSLImplementation > > And results with/without my changes are the same. > > I'm testing with original Tomcat 9.0.0.M9 and Tomcat 9.0.0.M9 + only my > > changes. > > I'm testing with blocking and non blocking. > > > > Just for testing purposes I changed > > org.apache.catalina.connector.OutputBuffer > > to use direct ByteBuffer and then executed SSL scenarios and again the > > results are the same. > > No visible performance improvement/degradation. > > > > Do you have some scenarios in mind in which a performance degradation can > > be experienced. > > I can test them and take a deeper look. > > > > Ok, good, normally it's the best thing to test. If using OpenSSL, you > should see better performance with direct buffers though, so maybe there's > something "wrong" with the testing. If you plan more ByteBuffer use, then > go ahead and I'll test it.
My idea was to introduce the changes step by step, because of this I wanted to backport the current changes to Tomcat 8.5 and receive a feedback. But if you propose to introduce all ByteBuffer related changes together, I'm OK and can commit the next changes. I saw several places where switch to ByteBuffer can be done and I'm going to finish and commit it. > > > > > > > - I didn't test again after the header processing refactoring (the old > > code > > > was there for speed back then in early 2000s, most likely it doesn't make > > > any difference now but it could be verified as well). > > > > I executed many tests and didn't see any problems. > > > > Yes, I had verified this one already :) I really appreciate your reviews and testing. Thanks, Violeta > Rémy
