------
They also how that proxies introduce additional bandwidth into the
network that often negates the benefits of sharing. Without a proxy
aborting a request interrupts document transfer, thus onsuming no more
bandwidth. However, ther is often a bandwidth mismatch between
connections between clients and proxies and proxies and servers. In
many cases the proxy has already downloaded much of a document before
it receives the abort notification. The exact results of this
mismatch depends on connections and how a proxy is set up to handle an
abort. In the experiments in this paper the total number of bytes
received by clients without a proxy before an abort was 49.8 GB. In
the worse case scenario, where a proxy continues to download after an
abort message, the total number of bytes received by clients was 58.7
GB, an 18% increase. In order to break even with the savings from
caching (reduction in bandwidth due to sharing), the proxy would have
to abort as soon as it received an abort message and the network bandwidth
would have to be pretty low (in these tests 0.5 Mbps when the client
proxy bandwidth was 21Kbps). And even in this best case scenario the
overall savings is reduced to just 6%.
In the end they found that the most beneficial scheme was to cache TCP
connections and not the actual page content. Basically, persistent
connections will reduce latency more than caching data.
---------
---------- Forwarded message ----------
Date: Mon, 02 Nov 1998 12:59:24 -0800 (PST)
From: Geoff Voelker <voelker@cs.washington.edu>
To: websys@cs
Subject: paper
The last paper Denise reviewed looks very interesting; we should all
read it. I'll bring copies to the meeting. Her review:
http://www.cs.washington.edu/research/networking/websys/local/summaries/caceres-proxy-devil-att98.tx
-geoff