Doesn't a TLS terminator proxy solve this? E.g. I usually put my application services behind HTTPS-enabled nginx and it works wonderfully.
Nope, so, the goal here is to reduce how many machines (each with their own RAM limits) are used. The task is holding open bidirectional SSL wrapped long-lived websocket connections. They're held open for hours at a time, since we need to send notifications when we get them.
Every connection has a base cost of the TCP kernel send/recv buffer, which in our case we dropped a bit to 4kb each. So that's still 8kb per connection right there. If we terminate the SSL on a separate machine from where we handle the connection, then it means we'll be using 8kb more memory per connection. Probably even greater because nginx has its own send/recv buffers for data.
I'm sure our use-case is a unique one, most people care about raw through-put so the majority of SSL optimization has focused on lowering CPU use under high load rather than memory use under massive amounts of connections.
What is your (unique) use-case about? What service do you provide to your users?
Do you have any insight about the mention in the slides of Google having a 10kb solution?
SSL is extremely expensive, on RAM (perhaps the implementations have optimized for throughput over RAM). I have yet to benchmark any SSL implementation in any language, with any binding, that can use less than 20kb per SSL connection. I mentioned in my talk here that SSL is very expensive, here's my benchmark suite that others may add to: https://github.com/bbangert/ssl-ram-testing/
I have implementations in several languages, so far both Go and Python 3.4 can get as low as the 20kb cited. If you can get your per-connection state below 20kb, then merely adding SSL means doubling or worse your RAM requirements, which is huge.
I appreciate that everyone loves obsessing on the language wars, but the SSL RAM overhead affects us regardless of language. I covered that in one of the slides near the end, it'd be great to see some movement on reducing the RAM footprint here.