According to OS X Instruments, 60% of time is spent writing the Date header

[reem]

This happens in src/response and is responsible for (according to profiles) 60+% of the running time of the server.

[seanmonstar]

Yea, I noticed the slowdown when I added that header. strftime is not too fast in libtime. I filed rust-lang/rust#18556 to try to fix this.

[reem]

In profiles it looks like most of the time is spent in rust_timegm, which is a C foreign function.

[seanmonstar]

Ah, would that be because of the tm.to_utc() call?

[reem]

Yes, I think.

[seanmonstar]

Yea, so to_utc calls Tm.to_timespec(), which is calling rust_timegm: http://doc.rust-lang.org/src/time/home/rustbuild/src/rust-buildbot/slave/nightly-linux/build/src/libtime/lib.rs.html#327

[reem]

I wonder if that will improve with runtime removal.

[seanmonstar]

It seems like it's a waste of calls. The Tm is already in UTC, since the default is calling time::utc_now(). It notices the gmt_offset is 0, so it seems like it could just return itself quickly without touching the C function...

[seanmonstar]

While that should be fixed in libtime, I could try checking for tm_gmtoff == 0 myself, and if so, no need to call to_utc().

[seanmonstar]

The TmFmt patch was merged 2 days ago. With that and the #100 patch, does this improve things noticeably?