Expand benchmarks about conscrypt HOT 10 CLOSED

No. We currently don't measure the impacts of ALPN. PRs welcome (Netty -> Conscrypt ftw) 😉

from conscrypt.

Running the benchmarks, I'm seeing a significant regression in our engine performance. Investigating.

from conscrypt.

A couple things that have already been done:

• Switched the default cipher to TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256 for all benchmarks
• Explicitly setting number of threads to 1 for all benchmarks

from conscrypt.

I'm not quite sure how we can benchmark the handshake without creating a new pair of engines each time, which would probably dominate the benchmark.

@Scottmitch any thoughts?

from conscrypt.

An engine is a socket, right? If there's a high setup cost outside the handshake in creating a socket, it seems probably fair to count that in a handshake benchmark.

from conscrypt.

@davidben no, an engine is not a socket ... creation is likely much lighter, but still not insignificant.

If there's a high setup cost outside the handshake in creating a socket, it seems probably fair to count that in a handshake benchmark.

Perhaps. I guess we can just try it out and see what the numbers look like.

from conscrypt.

Er, sorry, by socket I mean whether it's a per-connection thing or something shared between connections. If it's shared between connections then, yeah, the cost of creating one shouldn't be in the benchmark since it'll be amortized over many connections. If it's one per connection, then the cost of creating one, along with the handshake, figures into setup cost.

from conscrypt.

@nmittler - See Netty's SslEngineHandshakeBenchmark for ideas. We currently count the engine creation in the benchmark but could extract this out if necessary. I don't recall measuring the impacts of moving the creation/destruction outside the benchmark.

from conscrypt.

@Scottmitch thanks for the link! Yeah, after doing some profiling it doesn't seem that creation has much of an impact.

from conscrypt.

@Scottmitch after #239 goes in, everything will be addressed except ALPN. Does Netty have a benchmark for this that I could borrow?

from conscrypt.