Transcript Web Proxy Caching: The Devil is in the Details

Web Proxy Caching: The
Devil is in the Details
Ramon Cacere
Fred Douglis
Anja Feldmann
Gideon Glass
Michael Rabinovich
AT&T Labs-Research
Florham Park, NJ, USA
Brief Review
clients
servers
Reply
Req.
proxy
Req.
Reply
Brief Review
Client send requests to the proxy.
If the requested document is in its cache, the
proxy serves the request from its cache.
Otherwise, the proxy forward the request to the
server.
Server replies the request through the proxy
(proxy keep a copy of the requested document).
How does proxy caching
improve performance?
Reduce the user-perceived latency
associated with obtaining Web
documents.
Lower the network traffic from the Web
servers.
Reduce the service demands on content
providers.
Previous Work
High level details: hit ratio & byte hit ratio
Ignored exceptional cases such as
connection aborts.
Omitted the effect of cookies on
cacheability of resources.
This paper argues that...
Low-level details have a strong impact on
performance, particularly in heterogeneous
bandwidth environments.
Aborted trasfers can contribute significantly to total
bandwidth requirements.
“Cookies” dramatically affect the cacheability of
resources; therefore, affect the latency.
Caching TCP connections at proxy can reduce latency
more simply caching data.
Simulation
Web proxy simulator (PROXIM)
Workload: trace from AT&T Worldnet
12 days dialup traffic on a FDDI ring
encrypted IP addresses
contained information on both TCP events
and HTTP events
Simulator: PROXIM
Simulator Cache
sufficiently large
included proxy overhead in the request service time
Network Connections
zero or more open connections (cache-to-proxy &
proxy-to-server)
Proxy closes client-to-proxy connections with 3
minutes of idle time.
Proxy-to-server connections are timeout after 30 secs
of idle time.
Simulator: PROXIM (cont.)
Document Transfer
Packet-level delivery with TCP slow-start
1500-byte packets
constant round-trip time estimate for each
connection
Latency Calculations
connection setup time
HTTP request-response overhead
document transfer time
Results: (Hit Ratio)
When taking cookies into account
Hit ratio decreases from 54.5% to 35.2%.
Byte hit ratio decreases from 40.9% to
30.42%.
Solution: Techniques aimed at enabling
caching documents with cookies are important
for increasing hit rate.
Results:
(Bandwidth Savings)
When the proxy is present, the bandwidth
consumption of aborted requests is higher
due to the bandwidth mismatch between
the connections of client-to-proxy and
proxy-to-server.
Question: how much would this be offset
by the savings from caching?
Results:
(Latency Reduction)
Caching has limited effect on improving
latency (reduced the mean by 3.4%, the
median by 4.2%)
Solution: Maintain persistent connections
between clients and servers
Proxy as a connection cache.
Re-use persistent proxy-to-server connection
for obtaining documents for multiple clients.
Questions:
How does the proxy manage a connection
cach?
How many simultaneous connections it
should maintain with a server or a client.
Conclusion:
For dialup users
Hit ratios is lower than those reported
previously.
Bandwidth savings non-exist or is negative.
Latency reduction coming mostly from
caching TCP connections rather than
documents.