-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi y'all, its weekly update time

* Index:
1) 0.4.1.1 status
2) Pretty pictures
3) 0.4.1.2 and 0.4.2
4) Bundled eepserver
5) ???

* 1) 0.4.1.1 status

After a pretty bumpy 0.4.1 release (and subsequent rapid 0.4.1.1
update), the net seems to be back to normal - 50-something peers
actrive at the moment, and both irc and eepsites(I2P Sites) are reachable.  Most
of the pain was caused by insufficient testing of the new transport
outside lab conditions (e.g. sockets breaking at strange times,
excessive delays, etc).  Next time we need to make changes at that
layer, we'll be sure to test it more widely prior to release.

* 2) Pretty pictures

Over the last few days there have been a large number of updates
going on in CVS, and one of the new things added was a new stat
logging component, allowing us to simply pull out the raw stat data
as its being generated, rather than deal with the crude averages
gathered on /stats.jsp.  With it, I've been monitoring a few key
stats on a few routers, and we're getting closer to tracking down the
remaining stability issues.  The raw stats are fairly bulky (a
20-hour run on duck's box generated almost 60MB of data), but thats
why we've got pretty pictures - http://dev.i2p.net/~jrandom/stats/

The Y axis on most of those is milliseconds, while the X axis is
seconds.  There are a few interesting things to note.  First,
client.sendAckTime.png is a pretty good approximation of a single
round trip delay, as the ack message is sent with the payload and
then returns the full path of the tunnel - as such, the vast majority
of the nearly 33,000 successful messages sent had a round trip time
under 10 seconds.  If we then review the client.sendsPerFailure.png
along side client.sendAttemptAverage.png, we see that the 563 failed
sends were almost all sent the maximum number of retries we allow (5
with a 10s soft timeout per try and 60s hard timeout) while most of
the other attempts succeeded on the first or second try.

Another interesting image is client.timeout.png which sheds much
doubt on a hypothesis I had - that the message send failures were
correlated with some sort of local congestion.  The plotted data
shows that the inbound bandwidth usage varied widely when failures
occurred, there were no consistent spikes in local send processing
time, and seemingly no pattern whatsoever with tunnel test latency.

The files dbResponseTime.png and dbResponseTime2.png are similar to
the client.sendAckTime.png, except they correspond to netDb messages
instead of end to end client messages.

The transport.sendMessageFailedLifetime.png shows how long we sit on
a message locally before failing it for some reason (for instance,
due to its expiration being reached or the peer it is targetting
being unreachable).  Some failures are unavoidable, but this image
shows a significant number failing right after the local send timeout
(10s).  There are a few things we can do to address this:
 - first, we can make the shitlist more adaptive- exponentially
increasing the period a peer is shitlisted for, rather than a flat 4
minutes each.  (this has already been committed to CVS)
 - second, we can preemptively fail messages when it looks like
they'd fail anyway.  To do this, we have each connection keep track
of its send rate and whenever a new message is added to its queue, if
the number of bytes already queued up divided by the send rate
exceeds the time left until expiration, fail the message immediately.
 We may also be able to use this metric when determining whether to
accept any more tunnel requests through a peer.

Anyway, on to the next pretty picture -
transport.sendProcessingTime.png.  In this you see that this
particular machine is rarely responsible for much lag - typically
10-100ms, though some spikes to 1s or more.

Each point plotted in the tunnel.participatingMessagesProcessed.png
represents how many messages were passed along a tunnel that router
participated in.  Combining this with the average message size gives
us an estimated network load that the peer takes on for other people.

The last image is the tunnel.testSuccessTime.png, showing how long it
takes to send a message out a tunnel and back home again through
another inbound tunnel, giving us an estimage of how good our tunnels
are.

Ok, thats enough pretty pictures for now.  You can generate the data
yourself with any release after 0.4.1.1-6 by setting the router
config property "stat.logFilters" to a comma seperated list of stat
names (grab the names from the /stats.jsp page).  That is dumped to
stats.log which you can process with
  java -cp lib/i2p.jar net.i2p.stat.StatLogFilter stat.log
which splits it up into seperate files for each stat, suitable for
loading into your favorite tool (e.g. gnuplot).

3) 0.4.1.2 and 0.4.2

There have been lots of updates since the 0.4.1.1 release (see the
history [1] for a full list), but no critical fixes yet.  We'll be
rolling them out in the next 0.4.1.2 patch release later this week
after some outstanding bugs relating to IP autodetection are
addressed.

The next major task at that point will be to hit 0.4.2, which is
currently slated [2] as a major revamp to the tunnel processing.  Its
going to be a lot of work, revising the encryption and message
processing as well as the tunnel pooling, but its pretty critical, as
an attacker could fairly easily mount some statistical attacks on the
tunnels right now (e.g. predecessor w/ random tunnel ordering or
netDb harvesting).

dm raised the question however as to whether it'd make sense to do
the streaming lib first (currently planned for the 0.4.3 release).
The benefit of that would be the network would become both more
reliable and have better throughput, encouraging other developers to
get hacking on client apps.  After that's in place, I could then
return to the tunnel revamp and address the (non-user-visible)
security issues.

Technically, the two tasks planned for 0.4.2 and 0.4.3 are
orthogonal, and they're both going to get done anyway, so there
doesn't seem to be much of a downside to switching those around.  I'm
inclined to agree with dm, and unless someone can come up with some
reasons to keep 0.4.2 as the tunnel update and 0.4.3 as the streaming
lib, we'll switch 'em.

[1] http://dev.i2p.net/cgi-bin/cvsweb.cgi/i2p/history.txt?rev=HEAD
[2] http://www.i2p.net/roadmap

* 4) Bundled eepserver

As was mentioned in the 0.4.1 release notes [3], we've bunded the
software and configuration necessary for running an eepsite(I2P Site) out of
the box - you can simply drop a file in the ./eepsite/docroot/
directory and share the I2P destination found on the router console.

A few people called me on my zeal for .war files though - most apps
unfortunately need a little more work than simply dropping a file in
the ./eepsite/webapps/ dir.  I've put together a brief tutorial [4]
on running the blojsom [5] blogging engine, and you can see what that
looks like on detonate's site [6].

[3] http://dev.i2p.net/pipermail/i2p/2004-September/000456.html
[4] http://www.i2p.net/howto_blojsom
[5] http://wiki.blojsom.com/wiki/display/blojsom/About+blojsom
[6] http://detonate.i2p/

* 5) ???

Thats about all I've got at the moment - swing on by the meeting in
90 minutes if you want to discuss things.

=jr

-----BEGIN PGP SIGNATURE-----
Version: PGP 8.1

iQA/AwUBQWL3MxpxS9rYd+OGEQLk1gCfeMpSoYfbIlPWobks3i7lr8MjwDkAoOMS
vkNuIUa6ZwkKMVJWhoZdWto4
=hCGS
-----END PGP SIGNATURE-----