Performance Comparison of Swift http-parser port

Source Repository is:

git@github.com:helje5/HTTPParser.git

Swift port was branched of C http-parser Version 2.6.0, git hash:

bee4817ebea0eaa592143c7825f5c96f040c84e4

Swift port version this started from:

d3912dd40b4c2db9d23932649404e6e58f7ba013

Port Specific Notes

The port has NOT been optimized for speed at all.

• macros have been replaced with func's
  • so the compiler has to be smart to inline them
  • the C macros have been accessing local state, to simulate that I used nested functions
• the big goto loop has been replaced with a step function,
• there are a few dispatch tables which are converted straight to Swift arrays, such are range-checked etc
• callbacks are done using regular Swift closures

Ideas to get the Swift faster

Instruments says that almost all the time is lost in ARC.

• the dispatch tables are stored in global-let's, does this always have a dispatch_once overhead?
  • the parser might need to grab a copy of the pointer on startup or execute, but the macros would need to move into the parser.
• the callback closure invocations are pretty slow, we may be able to convert such to convention(c) closures
  • Update: this does not seem to be possible at the moment as Swift structs can't be marked as convention(c) (hence the callbacks can't be either)
    • there is an unofficial convention(thin), but that is too easy to crash
    • we may be forced to switch to a delegate protocol to get appropriate performance
• would unwrapping the nested funcs have any perf advantage?

RESULTS

Bottom line so far:

• C version is very fast (>20m requests per sec)
• Swift 2.3 is 3x slower than C, but 6.5x faster than Swift 3p4
• Swift 3p4 is really slow, maybe debug code related?

Obviously I may have made mistakes, as before :-)

2016-08-08 master

MacPro 3.7 GHz Quad-Core Intel Xeon E5

C Version - Release Build (2016-08-08):

[0] time: 93ms
[1] time: 93ms
[2] time: 96ms
[3] time: 95ms
[4] time: 94ms
-----
Total: 471ms COUNT: 10002000

Swift 2.3 (2016-08-08) - Release Build, safety checks off

struct vtable:

[0] time: 311ms
[1] time: 306ms
[2] time: 308ms
[3] time: 305ms
[4] time: 308ms
-----
Total: 1538ms COUNT: 10002000

protocol with closures:

[0] time: 1919ms
[1] time: 1953ms
[2] time: 1934ms
[3] time: 1916ms
[4] time: 1932ms
-----
Total: 9654ms COUNT: 20004000

Swift 3 Preview 4 (2016-08-08) - Release Build, safety checks off

struct vtable:

[0] time: 2021ms
[1] time: 2017ms
[2] time: 2027ms
[3] time: 2020ms
[4] time: 2021ms
-----
Total: 10106ms COUNT: 10002000

protocol with closures:

[0] time: 3865ms
[1] time: 3873ms
[2] time: 3905ms
[3] time: 3875ms
[4] time: 3880ms
-----
Total: 19398ms COUNT: 20004000

Instruments Says

• lots of ARC killing perf
  • TODO: check with latest w/o closures

struct-parser branch

• this is in master now

lookup-tables branch

• this is in master now

always-inline branch

• this is in master now