2015-05-01 19:28 GMT+02:00 Mark Thomas <ma...@apache.org>: > On 01/05/2015 15:03, Rémy Maucherat wrote: > > 2015-05-01 11:55 GMT+02:00 Mark Thomas <ma...@apache.org>: > > >> So, should we drop one of NIO or NIO2 in Tomcat 9? If not, why not? If > >> yes, which one? > >> > >> I volunteer to help maintain NIO2 for the time being, so I am not in > favor > > of removing it at the moment. Unfortunately, I think all three connectors > > have some good points. > > > > APR: > > - Blocking IO > > - OpenSSL (for the time being) > > If you can get OpenSSL working with the Java connectors then that opens > up the question why keep the APR/native connector. But we aren't there > yet. > > > - Sendfile > > > > NIO: > > - It's been around longer :) > > - Sendfile > > > > NIO2: > > - Modern async IO > > - Scatter / gather IO that can be exposed and taken advantage (see the > new > > IO calls I added; implementing them with APR and NIO is going to be a > whole > > lot more convoluted ...) > > - Probably HTTP/2 and Servlet.next will take advantage of it just like > > websockets did > > WebSockets didn't take advantage of it scatter/gather. Neither did it > take advantage of the async style of API. >
Websockets could easily take advantage of it. For example the code in WsRemoteEndpointImplServer and all its for loops over arrays of buffers to write them looks rather close to a gather write to me and could probably all be replaced with a single write. > > I don't think NIO has gotten any better, it's still the most horrendous > IO > > API imaginable as far as I am concerned. Of course, you can use > frameworks > > and stuff but ... I agree to keep it as well, since as you say it's more > > mature and stable, but that's about it. > > The main driver for this thinking is reducing complexity in the > connectors. Having to support both poller style and async style basic > I/O creates complexity. If we only supported one style the I/O code > could be a lot cleaner. > Of course, things can always be cleaner, but this causes issues. For example, the one I dislike the most is NIO, but as you said it is the most stable, so removing it is probably not a good idea. > > Overall, I think I prefer the async style. It is much easier to simulate > blocking and requires less supporting code (no pollers etc). > > But, APR uses the poller style and it has much better SSL performance. > > Thinking ahead (not sure how far, maybe Tomcat 9 at the outside) but if > we had OpenSSL working with NIO.2 and performance was similar to the > APR/native connector I'd be all for dropping NIO and APR/native in > favour of NIO2 and with the option to use OpenSSL or JSSE for SSL. > I was planning to send an email about that on Monday, so I guess I'm sending it now instead. > > Is that too extreme? > > The list of "pros" for each connector is not empty, so I don't see why it helps to remove one. Rémy
