[Cross-posted to dev.platform:  followup to dev.tech.network]

We've got a proposal and now a patch to implement resource packages:

   http://limi.net/articles/resource-packages/
   https://bugzilla.mozilla.org/show_bug.cgi?id=529208

Now we need to figure out if we want to take the patch or not.
Various people like or dislike the idea.   Feel free to chime in with
your opinion.

Meanwhile we'd like to have some numbers to help in our decision-
making.  We have some preliminary performance results:

  https://bugzilla.mozilla.org/attachment.cgi?id=455820

Here's some comments on them:

The results are promising.  But there are a few more tests that would
be really nice to see:

1) Your test has 126 resources for 1.3 MB total bandwidth (i.e. like
cnn.com).  It'd be nice to have at least one or two more data points
with fewer/smaller pages (like the bbc or amazon pages, with ~80
resources for 300-400KB total).   We don't care much about the low end
(unlikely to use resource packages), but I do think we care about the
mid-range.

2) The lowest latency you test is 55ms.  I'd like to see lower
latencies tested.  At home on my cable connection I see only 15ms ping
to foxnews.com (full disclosure: I tried nytimes.com and cnn.com
first, but they don't reply to pings :).   So far your numbers are
good at all latencies, provided you've got decent bandwidth.  I'd like
to know if that's true across the board.

3) Mobile latencies/bw:  this site

   http://www.yuiblog.com/blog/2010/04/08/analyzing-bandwidth-and-latency/

sees averages of 400ms/440kbps, for which you'd see a nice 33% speedup
(7.9 vs 12.4 seconds) for your big page.  But I'm wondering about lost
packets.  Resource pkgs are using one TCP connection, versus our
default of using 6 persistent HTTP connections, which are likely to
handle dropped packets better (the other 5 connections slurp up the b/
w while the lost packet cxn waits to retransmit).  So I'd like to see
some tests--including histograms, not just average/median times--over
mobile and/or some other lossy network.  (Maybe you could drag a
laptop to the edge of our wireless network?).

4) You mentioned at the Summit that it would be possible for a proxy
server (like squid) to dynamically construct a resource package "on
the fly" as a page is requested.  If we could knock off such a proxy
(perhaps just in a scripting language?  Apache module?), and have a
bunch of people use it for a while on different networks, we could get
a lot of interesting numbers.   My thought here is that the proxy
would randomly choose to reply with the resource package or not, and
we could keep stats on the average/variance timings to see how
resource packages compare with vanilla proxying (we could always
construct the package and delay replying until it's done, even if
we're not using it, to keep the pkg construction costs out of the
picture).  A side benefit here is that we might wind up with an
accelerator (or at least a design for one that could be written in C/+
+) that might be the easiest way to start rolling out widespread use
of resource packages on the intarwebs.

Jason
_______________________________________________
dev-tech-network mailing list
[email protected]
https://lists.mozilla.org/listinfo/dev-tech-network

Reply via email to