Re: Some data related to the frequency of cache-busting

Benjamin Franz points out several ways in which my simplistic trace
analysis might have overestimated the number of possibly cache-busted
responses seen at our proxy.

In particular, he suggests that some of the non-query
possibly-cachable references that I counted might actually
have been CGI output, which should not have been included
in the set of "possibly cache-busted responses".  (I will
note, however, that one of the examples he gave would NOT
have been counted as such by my analysis, because that URL
included the string "cgi-bin".  I explicitly did not count
such URLs.)

If someone would like to propose a *feasible* filter on URLs
and/or response headers (i.e., something that I could implement
in a few dozen lines of C) that would exclude other CGI
output (i.e., besides URLs containing "?" or "cgi-bin", which
I already exclude), then I am happy to re-run my analysis.

Drazen Kacar pointed out that I should probably have
excluded .shtml URLs from this category, as well, because
they are essentially the same thing as CGI output.  I checked
and found that 354 of the references in the trace were to .shtml
URLs, and hence 10075, instead of 10429, of the references
should have been categorized as possibly cache-busted.  (This
is a net change of less than 4%.)

    I would say the only *confirmable* deliberate cache busting done
    are the 28 pre-expired responses. And they are an insignificant
    (almost unmeasurable) percentage of the responses.

If I was writing a scientific paper whose thesis was that a
significant fraction of the responses are cache-busted, then
you are right that I would not have a rigorous proof regarding
anything but these 28 pre-expired responses.  And, no matter
how much more filtering I do on the data, I would not expect
to be able to construct a rigorous proof based on such a trace.

On the other hand, I don't believe that this trace could provide
a rigorous proof of the converse hypothesis, that no deliberate
cache-busting is done.  Nor do I believe that any trace-based
analysis could prove this, given the frequency with which I
found responses that leave the question ambiguous.

In short, if we are looking for a rigorous, scientific *proof*
that cache-busting is either prevalent or negligible, I don't
think we are going to find it in traces, and I can't think of
where else one might look.

But we are engaged in what fundamentally is an *engineering*
process, rather than a scientific one.  This means that, from
time to time, we are going to have to infer future reality from
an imprecise view of current reality, and that the future is
in large part determined by the result of our engineering, not
independent of it.

I welcome other sources of data that might help make this inference
more reliable.  Certainly we should not base everything on five
hours of trace data from one site.  On the other hand, it's
foolish to dismiss the implications of the data simply because
it fails to rigorously prove a particular hypothesis (pace the
Tobacco Institute, which has taken about 30 years to admit that
there might in fact be a connection between smoking and cancer.)

    As you noted - much more study is needed. This one is utterly
    inconclusive. You conclude from your numbers that significant
    savings can be found.

I wouldn't say I concluded that.  I said "there does seem to
be some potential here."

    I conclude from the same numbers that the extra overhead of the hit
    metering in fact is *higher* than the loses to deliberate cache
    busting. You would have more network traffic querying for hit meter
    results than the savings for such a tiny number of cache busted

This mystifies me.  What overhead of hit-metering are you talking about?

There are three kinds of overhead in our proposed scheme:

	(1) additional bytes of request headers
		(a) for agreeing to hit-meter
		(b) for reporting usage-counts
	(2) additional bytes of response headers
	(3) additional HEAD request/response transactions for
		"final reports"

Overheads of types #1(b), #2, and #3 are *only* invoked if the origin
server wants a response to be hit-metered (or usage-limited,
but that's not relevant to this analysis).  This means that
if hit-metering were not useful to the origin-server, it would
not be requested, and so these overheads would not be seen.
(I'm assuming a semi-rational configuration of the server!)

Note that #3 can *only* happen instead of a full request
on the resource, and is likely to elicit a smaller (no-body)
response, so it's not really clear that this should be
counted as an "overhead".

What remains is the overhead (type #1(a)) of a proxy telling
a server that it is willing to meter.  I'll ignore the obvious
choice that a proxy owner could make, which is to disable this
function if statistics showed that hit-metering increases overheads
in reality, and assume that the proxy is run by someone of less
than complete understanding of the tradeoffs.

So, once per connection, the proxy would send
	Connection: meter
which is 19 bytes, by my count.  If each connection carried just
one request, then (assuming that the mean request size stays
at about 309 bytes, which is what I found for all of the requests
I traced, and this does not include any IP or TCP headers!), then
this is about a 6% overhead. (But at one request/connection,
and with a mean request size smaller than 576 bytes, there would
probably be almost no increase in packet count.)

However, since hit-metering can only be used with HTTP/1.1 or
higher, and persistent connections are the default in HTTP/1.1,
and because we defined this aspect of a connection to be "sticky"
in our proposal, one has to divide the calculated overhead by
the expected number of requests per connection.  As far as I know,
nobody has done any quantitative study of this since my SIGCOMM '95
paper, which is presumably somewhat out of date, but (using simulations
based on traces of real servers) I was expecting on the order of 10
requests/connection.  It might even be higher, given the growing
tendency to scatter little bits of pixels throughout every web page.

Anyway, I wouldn't presume to put a specific number on this, because
I'm already basing things on several layers of speculation.  But I
would appreciate seeing an analysis based on real data that supports
your contention, that "the extra overhead of the hit metering in fact
is *higher* than the loses to deliberate cache busting."


Received on Monday, 2 December 1996 19:58:21 UTC