Communications - December 2012

page-load time across browser instantiations using SPDY, 13 though a caveat was that domain names in the study were neither recorded nor weighted. Google itself is thought by Google developers to be the current dominant consumer of server-side SPDY technology so is likely overrepresented in these results. Google sites were, in 2011, already heavily optimized, suggesting the stated improvement was likely conservative, though further data is needed for confirmation.

Google’s second result from the study was that (encrypted) SPDY is faster than (unencrypted) HTTP for Google’s AJAX search; Google researchers provided no further detail.

Google lab tests set one. Google performed a series of laboratory benchmarks of SPDY vs. HTTP under various conditions, though unencrypted SPDY, which would be expected to be faster than encrypted SPDY, was compared against HTTP, despite SPDY deployments being predominantly encrypted.

For simulated downloads of the top 45 pages on the Web (as recorded by Alexa), Google in 2011 reported a 51% reduction in uploaded bytes, 4% reduction in downloaded bytes, and 19% reduction in total packets vs. HTTP. 13 Uploaded bytes were significantly reduced due to SPDY’s header compression and the fact that HTTP headers are amenable to strong compression. Google reported that download bytes were only marginally reduced, as most downloaded content in its tests was not headers and in many cases already compressed. The reduction in total packets is due to both a reduction in overall bytes and the fact that SPDY uses only a single connection, resulting in more “full” packets.

Google lab tests set two. A 2009 Google SPDY white paper14 described simulated page-load time of SPDY vs. HTTP for the Alexa top 25 Web sites. The first test simulated SPDY vs. HTTP with 1% packet loss over simulated home-network connections. Unencrypted SPDY exhibited 27%–60% improvement in page-load time, and encrypted (SSL) SPDY exhibited 39%–55% improvement in page-load time. A second test determined how packet-loss affected SPDY (unencrypted) vs. HTTP; at 0% packet loss, SPDY was 11% faster, and at 2%

figure 7. httP request and response with example headers; header keys (such as “user-agent”) and header values (such as a particular user agent string) are repeated many times over on a typical connection so make good candidates for compression.

GET /pub/ WW W/ picture.jpg HT TP/1.1

Host: www.w3.org …

Accept-Endcoding: gzip, deflate, sdch

User-agent: Mozilla/5.0 (compatible; MSIE 9.0; Windows NT 6. 1; WOW64; Trident/5.0)

HTTP/1.1 200 OK
Date: Thu, 26 Jan 2012 04:16: 13 GM T
…
Server: Apache/2.2.3 (Red Hat)
Content-Type: image/jpeg

figure 8. httP is unable to push resources to the client, even when it knows the client will require them soon (left); sPDY server push allows the server to initiate the transfer of a resource it believes the client will need soon (right).

Standard HTTP
SPDY Server
client
REQexample.com/home
RES <html> … <img
src=" banner.jpg" /> …
REQexample.com/banner.jpg
RES [ banner.jpg]
server client
REQexample.com/home
PUSH [ banner.jpg]
server
Client
learns
it needs
banner.jpg
only after
parsing
html
RES <html> … <img
src=" banner.jpg" /> … </html>
Server
preempts
client’s
imminent
request for
banner.jpg
and pushes it,
removing
a round trip

packet loss SPDY was up to 47% faster. A third test simulated SPDY vs. HTTP as a function of round-trip time; for example, for a 20ms round trip, SPDY was 12% faster than HTTP, and for a 200ms round trip, SPDY was 26% faster; for a full exposition of these results, see Google. 14

Questions on SPDY performance.
There is a conspicuous lack of results
describing how SPDY performs on
mobile devices. SPDY’s dominant per-
formance improvements are due in
theory to reduced round trips between
client and server. Many mobile-carrier
technologies today exhibit latency sev-
eral times that of their fixed-line and
Wi-Fi counterparts. By some projec-
tions, the majority of the developing
world will have its first Internet experi-
ence through a mobile carrier, prolif-
erating high-latency connections. In
theory, SPDY is ideally suited to these
high-latency mobile environments,
though real-world results are needed
for confirmation.