On Wed, 29 Jun 2016, Tim Rühsen wrote:
I recently made a few comparisons between curl 7.50.0-DEV and wget 1.18 and
was astonished about wget outperforming curl by some fair amount on single
HTTPS request/response cycles.
They really *should* perform very similarly.
So my question goes... what is 'wrong' with that version of curl. Or what
did I oversee - maybe some special options ?
No special option should be necessary. Other than of course protocol version
differences and cipher selections etc.
Intel(R) Core(TM) i5-3570K CPU @ 3.40GHz
2MBit/s DSL, ping to www.google.com is ~106ms
Downloading a non-existent page at www.google.com (a.html) via HTTPS/HTTP1.1.
The figures from 'time' are the fastest I got in 10 tries.
$ time curl -s -o/dev/null https://www.google.com/a.html
$ time wget -q -o/dev/null --no-alpn https://www.google.com/a.html
I assume you meant --no-alpn on the curl command line?
I have 100mbit and 2.5ms ping time to www.google.com. I doubt it matters much
but my CPU is a Intel(R) Core(TM) i7-3770K CPU @ 3.50GHz
My local tests on your command lines:
[current git master using OpenSSL/1.1.0]
$ time ./src/curl -s -o/dev/null --no-alpn https://www.google.com/a.html
real 0m0.041s
user 0m0.020s
sys 0m0.004s
[1.18 from Debian, using GnuTLS)
$ time wget -q -o/dev/null https://www.google.com/a.html
real 0m0.044s
user 0m0.024s
sys 0m0.000s
[my Debian curl 7.47.0 using GnuTLS/3.4.13 I get much worse results]
$ time curl -s -o/dev/null --no-alpn https://www.google.com/a.html
real 0m0.159s
user 0m0.084s
sys 0m0.012s
I didn't check the TLS negotiated TLS ciphers.
Looks like there is some (pretty huge ~ 240ms) startup penalty hidden
somewhere. Building with OpenSSL makes not much difference.
Interesting find. Thanks for pointing this out. Clearly there are reasons to
take a deeper look into this.
--
/ daniel.haxx.se
-------------------------------------------------------------------
List admin: https://cool.haxx.se/list/listinfo/curl-library
Etiquette: https://curl.haxx.se/mail/etiquette.html
