[tls] write data by small (1500 bytes) chunks, see http://www.belshe.com/ #961
Conversation
benchmark? |
Sorry, no benchmarks. That is optimization for server client's - read article carefully! (Last time I'd posted that pull request - I missed that fact) |
Basically, if you're sending large chunks - browser can display parts of page before it stop receiving, b/c part of chunk can't be encrypted. |
BTW, try this in Chrome: It's nodejs SPDY server :) And it's sending data in 1500 bytes chunks |
I'm concerned about the effect this might have on throughput, and thats what it would be nice to have a benchmark of -- showing the effect is negligible. 1500 byte writes to SSL_write also would cause a bad behavior for many connections, because that would become 1500+ bytes with the TLS headers/protocol, and for most MTUs would end up being 2 packets instead of 1. (good thing to look at in wireshark over non-localhost) If the benchmark shows this isn't a huge drop in throughput, I'd be interested in seeing what happens with using ~1300-1400 byte writes, which should drop you bellow the common MTUs. |
Hm... Nice point, @pquerna. Mike has told me to use 1400 bytes, but I'd seen 1500 in chromium's flip_server - so I decided to use last one... Can you please help and perform some tests? |
I think we should come back to this issue, splitting can't be done in userland |
Reconsider? |
Can one of the admins verify this patch? |
fuck it. |
@indutny you should be more friendly to pull request submitters ;P |
Is it a special command for jenkins? |
This pull request was previously posted here, but this time it's low-level:
P.S.
See http://www.belshe.com/2010/12/17/performance-and-the-tls-record-size/