You're right - the mass resend is consistently overloading the connection. Logging client acks shows that the resent
packet bursts are not acked. Raising the default RTO only delays the resend since the client doesn't send acks no
matter how long the wait.
Setting enable_adaptive_throttles = true makes the problem go away for me - the client now receives the object update
packets at a rate that it can cope with. Is there really any reason not to turn this on by default? I hear what you're
saying about potential issues but it seems far better to get real world data to iron them out than persist with our
current naive implementation. At the worst we can always just disable it again.
Lowering the bw in the client settings also works, though it still allows a noticeable overload at the beginning of the
client connection. I'm very surprised that the throttling isn't being properly adjusted automatically, or is this what
enable_adaptive_throttles does (I traced it into the AdaptiveTokenBucket but no further yet)?
On 12/10/11 16:32, Mic Bowman wrote:
Couple things...
First... here's the link to the thread on some of the original issues... not
sure if this went in to dan's
retransmission fix but i'll get to that in a minute:
https://lists.berlios.de/pipermail/opensim-dev/2011-March/010029.html
Second... In spite of all the questions/comments below that this is the "expected"
behavior... there may be a bug & we
need more data to find it.
Question... do you have the adaptive throttling turned on? Look for
enable_adaptive_throttles. It is turned off by
default. That is the switch that is designed to fix the problem you are
seeing... but the code has some other behaviors
which are correct but... "different" than what most admins expect (the throttle
mimics the TCP slow start to identify
and avoid congestion though not quite as conservatively as the TCP algorithm).
One explanation for what you are seeing is that the network simply cannot send
as many packets as you are asking. This
is frequently true for long-haul connections. When we did our analysis, we saw a
"cliff". Once you pass the point where
the number of outgoing packets exceeds the physical networks ability to deliver
them ACKs are not processed in time and
very bad things happen. That is... enough ACKs are coming late enough so that
the resend packet is already in the queue.
The current version of the code handles that situation far better than the old
version (the current version is bound by
the number of dynamic objects... the old version would hold *every* update to
that object... even ones that were
superceded by later updates).
The only real solution to this problem is to stop sending updates so fast or to
be much more conservative about resends.
The first can be addressed one of three ways... First, configure the simulator
so that the outgoing bw per connection is
throttled much lower. Second, turn down the BW setting on the viewer. Third,
turn on the adaptive throttling code in the
simulator.
Being conservative about resends has its own share of problems... Basically
what this means is that you wait longer to
resend a packet. However... if you wait too long to retransmit, then a portion
of your scene is not updating. The
algorithm for determining when to retransmit is based on a commonly used
algorithm described in one of the RFCs (the
actual RFC is referenced in a comment in the opensim code).
As to the UDP packet vs going back through the SceneEntityUpdate queue... well
that one is a no brainer... If you use
the UDP packet for retranmission you are almost always sending useless old data
and just filling up the network with
fluff. Outside of the initial scene load, you see streams of updates for
objects more often than a single update. If you
retransmit the UDP packet, you are almost always sending old data. Going
through the sceneentityupdate queue basically
says "retransmit the current state of the object".
--mic
On Wed, Oct 12, 2011 at 5:56 AM, Justin Clark-Casey <[email protected]
<mailto:[email protected]>> wrote:
Yes, I may not have the correct cause in my original e-mail but I'm
certainly seeing a lot of resends that never
appear to stop. The effect is not consistent - sometimes I can log in to
the remote sim and there will be no
continuous resend. But most of the time a large proportion of updates
appear to be continuously resent.
I'm surprised no-one else has reported it though I did see a lot of
continuous updating on Wright Plaza @ osgrid.org
<http://osgrid.org> last night, which I don't think is related to changes
made by scripts. Perhaps it's only really
noticeable if you start looking, have visual object updates
development/advanced settings on in a viewer or are
wondering why viewer received packets keeps spiking.
As far as I can see from LLClientView/LLUDPServer, both prim and property
updates resends are happening but I need
to study the code more closely and possibly old e-mails.
I hope to do some more analysis later on today and be around on IRC.
On 12/10/11 09:20, Lake, Dan wrote:
Thanks for the additional info on this, Melanie. You are correct that
the resend should get a new sequence
number and that old updates for an object will no longer get sent when
there is a newer update to send.
Properties and updates are handled almost the same but different code.
I'll look it over with Mic in the morning
so I understand what is going on.
~Dan
-----Original Message-----
From: opensim-dev-bounces@lists.__berlios.de
<mailto:[email protected]>
[mailto:opensim-dev-bounces@__lists.berlios.de
<mailto:[email protected]>] On Behalf Of Melanie
Sent: Tuesday, October 11, 2011 11:22 PM
To: [email protected] <mailto:[email protected]>
Subject: Re: [Opensim-dev] Continual object update resends if acks are
missed
It appears this mechanism has already been removed for prim updates
because intermediate updates are needed fro proper motion of
physical prims, but was left in for property updates. However,
culling of later updates as well as preservation of sequence numbers
was not done.
Melanie
On 12/10/2011 07:43, Mic Bowman wrote:
We should collect more information on what is actually happening.
The old
behavior was to resend infinitely (and very badly). It would send
old
updates even when there were new updates for the same object. We
changed the
resend to transmit the current data rather than old data. If there
are still
no acks coming back after some time then we need to figure out what
the
viewer is expecting& when it no longer acks packets.
There is a write up on the procedure in an old email message on
this list.
--mic
On Tue, Oct 11, 2011 at 3:55 PM, Justin Clark-Casey<
[email protected] <mailto:[email protected]>> wrote:
Hi dslake (since this is chiefly addressed to you :)
In commit b5ab33b5 back on Wednesday 20th April 2011, you made
a change so
that the resend of object updates or property updates is
threaded back
through Resend methods on LLClientView rather than the normal
resend within
UDPServer.
The normal resend only performs the resend once, while going
through the
LLClientView.Resent*() methods will continually attempt the
resend until
acked as they put the requests back on the
m_entityUpdates/m_entityProps
queue.
From my experience, often the client will not reack such
packets. This
means that they are continuously resent until the client logs
out. I can
see this happening by uncommenting the log lines in
LLClientView.**__ResendPrimUpdates()
and ResentPropertyUpdates(). This is chiefly seen on remote
regions (I
can't repro on a localhost).
What do you think? Can we resend such packets just once?
Thanks,
--
Justin Clark-Casey (justincc)
http://justincc.org/blog
http://twitter.com/justincc
________________________________**_________________
Opensim-dev mailing list
[email protected]
<mailto:[email protected]>
https://lists.berlios.de/**__mailman/listinfo/opensim-dev
<https://lists.berlios.de/**mailman/listinfo/opensim-dev><h__ttps://lists.berlios.de/__mailman/listinfo/opensim-dev
<https://lists.berlios.de/mailman/listinfo/opensim-dev>>
_________________________________________________
Opensim-dev mailing list
[email protected] <mailto:[email protected]>
https://lists.berlios.de/__mailman/listinfo/opensim-dev
<https://lists.berlios.de/mailman/listinfo/opensim-dev>
_________________________________________________
Opensim-dev mailing list
[email protected] <mailto:[email protected]>
https://lists.berlios.de/__mailman/listinfo/opensim-dev
<https://lists.berlios.de/mailman/listinfo/opensim-dev>
_________________________________________________
Opensim-dev mailing list
[email protected] <mailto:[email protected]>
https://lists.berlios.de/__mailman/listinfo/opensim-dev
<https://lists.berlios.de/mailman/listinfo/opensim-dev>
--
Justin Clark-Casey (justincc)
http://justincc.org/blog
http://twitter.com/justincc
_________________________________________________
Opensim-dev mailing list
[email protected] <mailto:[email protected]>
https://lists.berlios.de/__mailman/listinfo/opensim-dev
<https://lists.berlios.de/mailman/listinfo/opensim-dev>
_______________________________________________
Opensim-dev mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/opensim-dev
--
Justin Clark-Casey (justincc)
http://justincc.org/blog
http://twitter.com/justincc
_______________________________________________
Opensim-dev mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/opensim-dev
