* increased the maximum number of fragments allowed in a message from 31 to 127,
reducing the maximum fragment size to 8KB and moving around some bits in the fragment
info. This is not backwards compatible.
* removed the old (hokey) congestion control description, replacing it with the TCP-esque
algorithm implemented
note: the code for the ACK bitfields and fragment info changes have not yet been
implemented, so the old version of this document describes whats going on in the live net.
the new bitfields / fragment info should be deployed in the next day or so (hopefully :)
* Cleaned up the peers page a bit more.
more udp stuff:
* add new config option: i2np.udp.alwaysPreferred=true to adjust the bidding
so that UDP is picked first, even if a TCP connection exists
* fixed the initial clock skew problem (duh)
* reduced the MTU to 576 (largest nearly-universally-safe, and allows a
tunnel message in 2 fragments)
* handle some races @ connection establishment (thanks duck!)
* if there are more ACKs than we can send in a packet, reschedule another
ACK immediately
* Added a small new page to the web console (/peers.jsp) which contains
the peer connection information. This will be cleaned up a lot more
before 0.6 is out, but its a start.
* Reduced some SimpleTimer churn
* add hooks for per-peer choking in the outbound message queue - if/when a
peer reaches their cwin, no further messages will enter the 'active' pool
until there are more bytes available. other messages waiting (either later
on in the same priority queue, or in the queues for other priorities) may
take that slot.
* when we have a message acked, release the acked size to the congestion
window (duh), rather than waiting for the second to expire and refill the
capacity.
* send packets in a volley explicitly, waiting until we can allocate the full
cwin size for that message
* more stats. including per-peer KBps (updated every second)
* improved blocking/timeout situations on the send queue
* added drop simulation hook
* provide logical RTO limits
* Reduce the peer profile stat coallesce overhead by inlining it with the
reorganize.
* Limit each transport to at most one address (any transport that requires
multiple entry points can include those alternatives in the address).
udp stuff:
* change the UDP transport's style from "udp" to "SSUv1"
* keep track of each peer's skew
* properly handle session reestablishment over an existing session, rather
than requiring both sides to expire first
* More fixes for the I2PTunnel "other" interface handling (thanks nelgin!)
* Add back the code to handle bids from multiple transports (though there
is still only one transport enabled by default)
* Adjust the router's queueing of outbound client messages when under
heavy load by running the preparatory job in the client's I2CP handler
thread, thereby blocking additional outbound messages when the router is
hosed.
* No need to validate or persist a netDb entry if we already have it
And for some udp stuff:
* only bid on what we know (duh)
* reduceed the queue size in the UDPSender itself, so that ACKs go
through more quickly, leaving the payload messages to queue up in
the outbound fragment scheduler
* rather than /= 2 on congestion, /= 2/3 (still AIMD, but less drastic)
* adjust the fragment selector so a wsiz throttle won't force extra
volleys
* mark congestion when it occurs, not after the message has been
ACKed
* when doing a round robin over the active messages, move on to the
next after a full volley, not after each packet (causing less "fair"
performance but better latency)
* reduced the lock contention in the inboundMessageFragments by
moving the ack and complete queues to the ACKSender and
MessageReceiver respectively (each of which have their own
threads)
* prefer new and existing UDP sessions to new TCP sessions, but
prefer existing TCP sessions to new UDP sessions
* Added button to router console for manual update checks.
* Fixed bug in configupdate.jsp that caused the proxy port to be updated
every time the form was submitted even if it hadn't changed.
* Added a pool of PRNGs using a different synchronization technique,
hopefully sufficient to work around IBM's PRNG bugs until we get our
own Fortuna.
* In the streaming lib, don't jack up the RTT on NACK, and have the window
size bound the not-yet-ready messages to the peer, not the unacked
message count (not sure yet whether this is worthwile).
* Many additions to the messageHistory log.
* Handle out of order tunnel fragment delivery (not an issue on the live
net with TCP, but critical with UDP).
* Added a pool of PRNGs using a different synchronization technique,
hopefully sufficient to work around IBM's PRNG bugs until we get our
own Fortuna.
* In the streaming lib, don't jack up the RTT on NACK, and have the window
size bound the not-yet-ready messages to the peer, not the unacked
message count (not sure yet whether this is worthwile).
* Many additions to the messageHistory log.
* Handle out of order tunnel fragment delivery (not an issue on the live
net with TCP, but critical with UDP).
and for udp stuff:
* implemented tcp-esque rto code in the udp transport
* make sure we don't ACK too many messages at once
* transmit fragments in a simple (nonrandom) order so that we can more easily
adjust timeouts/etc.
* let the active outbound pool grow dynamically if there are outbound slots to
spare
* use a simple decaying bloom filter at the UDP level to drop duplicate resent
packets.
And this is the big "Fix the Parser" patch. It turns the sam_parse function in src/parse.c into something that actually works. Generating the argument list from an incoming SAM thingy is a bit memory churn-y; perhaps when I have time I'll replace all those strdups with structures that simply track the (start,end) indices.
Oh and also I moved i2p-ping to the new system. Which required 0 change in code. All I did was fix the Makefile, and add shared library libtool support. Anyway, so enjoy folks. It's rare I'm this productive
- polecat
* In the SDK, we don't actually need to block when we're sending a message
as BestEffort (and these days, we're always sending BestEffort).
* Pass out client messages in fewer (larger) steps.
* Have the InNetMessagePool short circuit dispatch requests.
* Have the message validator take into account expiration to cut down on
false positives at high transfer rates.
* Allow configuration of the probabalistic window size growth rate in the
streaming lib's slow start and congestion avoidance phases, and default
them to a more conservative value (2), rather than the previous value
(1).
* Reduce the ack delay in the streaming lib to 500ms
* Honor choke requests in the streaming lib (only affects those getting
insanely high transfer rates)
* Let the user specify an interface besides 127.0.0.1 or 0.0.0.0 on the
I2PTunnel client page (thanks maestro^!)
(plus minor udp tweaks)
* Added the possibility for i2ptunnel client and httpclient instances to
have their own i2p session (and hence, destination and tunnels). By
default, tunnels are shared, but that can be changed on the web
interface or with the sharedClient config option in i2ptunnel.config.
2005-04-17 jrandom
* Marked the net.i2p.i2ptunnel.TunnelManager as deprecated. Anyone use
this? If not, I want to drop it (lots of tiny details with lots of
duplicated semantics).
* Added new user-editable eepproxy error page templates.
2005-04-17 jrandom
* Revamp the tunnel building throttles, fixing a situation where the
rebuild may not recover, and defaulting it to unthrottled (users with
slow CPUs may want to set "router.tunnel.shouldThrottle=true" in their
advanced router config)
* use the new raw i2np message format (the previous corruptions were due to above)
* add a new test component (UDPFlooder) which floods all peers at the rate desired
* packet munging fix for highly fragmented messages
* include basic slow start code
* fixed the UDP peer rate refilling
* cleaned up some nextSend scheduling
* Make sure we don't get cached updates (thanks smeghead!)
* Clear out the callback for the TestJob after it passes (only affects the
job timing accounting)
* include a new signedOnTime so that we can prepare the packet at a different moment from
when we encrypt & send it (also allowing us to reuse that signature on resends for the same
establishment)
* Security improvements to TrustedUpdate: signing and verification of the
version string along with the data payload for signed update files
(consequently the positions of the DSA signature and version string fields
have been swapped in the spec for the update file's header); router will
no longer perform a trusted update if the signed update's version is lower
than or equal to the currently running router's version.
* Added two new CLI commands to TrustedUpdate: showversion, verifyupdate.
* Extended TrustedUpdate public API for use by third party applications.
