1.
So that
was
HTTP/2,
what's
next?
February 4 2017

2.
Questions?
I’ll hang aroundI’ll hang around
outside theoutside the
room afterwardroom afterward

3.
Daniel Stenberg
@bagder

4.
Request and payload growth
in the last 5 years...
8080
100100
800K800K
2.5M2.5M
the average website loads 50+ resources on a
single domain

5.
Head of line blockingHead of line blocking

6.
Multiplexed (1/2)
Multiple streams over a single physical connection
Max number of parallel streams set by peer

7.
Multiplexed (2/2)

8.
Better utilization of bandwidth

9.
So that was HTTP/2
February 4, 2017
Adoption,
usage, support

10.
servers

11.
browsers

12.
Browser stats – Feb 2017
HTTP/2:
Share of HTTPS:
30%30%
60%60%

13.
Internetz – February 2017
12% of top 10 million
23% of top 10,000
26% of top 1000
>50% for your site

14.
Winners and
losers
So that was HTTP/2
February 4, 2017

15.
Did HTTP/2 meet its promises

16.
The remote corners of Internet
Percentile Desktop Mobile
5 1 11
25 20 44
50 79 94
75 194 184
95 800 913
Milliseconds RTT

17.
Queuing time h1 vs h2
(Time waiting internally to send off a HTTP request)
Percentile HTTP 1 HTTP 2
80 100 ms 2 ms
95 2000 ms 16 ms
>100ms: H1 20%, H2 3%

18.
0% packet loss
Image and data by Hooman Beheshti, Fastly

19.
2% packet loss
Image and data by Hooman Beheshti, Fastly

20.
Packet loss, hey?
Single connection vs 6
Head of line blocking
A single dropped packet
blocks everything

21.
Packet loss, hey?
TLS TLS TLS
HTTP/2
frame
HTTP/2
frame
HTTP/2
frame
HTTP/2
frame
HTTP/2
frame
IP IP IP IP IP IP
TCP TCP TCP TCP

22.
Packet loss, hey?
IP
HTTP/2
frame
HTTP/2
frame
HTTP/2
frame
HTTP/2
frame
HTTP/2
frame
TCP TCP TCP TCP
TLS TLS TLS
IP IP IP IP IP IP

23.
Fixing TCP head of blocking

24.
A non-blocking TCP + TLS + HTTP/2
Needs independent packets
… that still are stream aware
Needs new retransmissions/ACKs
New protocol?
Fixing TCP takes decades –
if even doable

25.
QUIC
“TCP”, TLS and HTTP/2 over UDP
no TCP head of line blocking
other congestion control
move across interfaces
“TCP improvements” - much faster
Google has this widely deployed already
UDP not as problematic as we thought

26.
Packet loss, hey?
IP
UDP UDP UDP UDP UDP UDP
quic quic quic quic quic quic
TLS TLS TLS TLS TLS TLS
h2 h2 h2 h2 h2 h2
IP IP IP IP IP IP

27.
The IETF QUIC wg
Just started
Massive interest
s/custom crypto/TLS 1.3
Will transfer more than HTTP/2 frames
First interim meeting in Tokyo in Jan ‘17
IETF-QUIC vs Google-QUIC
Maybe early live test around mid-2017
That meansThat means
maybemaybe andand
earlyearly teststests

28.
Neither 3 nor 2
Perhaps “HTTP/3”, but not in name
Perhaps “TCP/2”, but not in name
Just QUIC

29.
Round-up
HTTP/1 is not optimal
HTTP/2 is binary and multipled
HTTP/2 is widely used
HTTP/2 makes sites faster
(IETF-)QUIC is coming
QUIC is HTTP/2 frames over UDP

30.
Thank you!
Questions outside!