[Opensim-dev] Behaviour of adaptive throttles under high load
Justin Clark-Casey
jjustincc at googlemail.com
Mon Dec 1 19:38:43 UTC 2014
I agree with you about not stabbing in the dark. That is why I spent an enormous amount of treasure on improving
pCampbot/libomv, creating other tools and adding more stat and real time change capabiltiies to allow us to reproduce
issues before making changes.
So it's unfortunate that in this case that I simply ran out of time to properly investigate the issue with adaptive
throttles being decimated by a string of ack waits expiring in the same second.
My major concern is that this could well be the single remaining blocker to decent performance at higher avatar numbers
than has previously been possible in core out-of-the-box OpenSimulator (excpet for using mysql instead of sqlite). It
won't matter that every other issue has been addressed - just this one is potentially enough to still screw up
performance. So I really want to address this before making the next release. I really don't want to tell people
having problems that they should try switching off adaptive - OpenSimulator should work out-of-the-box for as broad a
selection of load profiles as possible.
The problem I face now is that in a perfect world one would definitely go back and do extensive testing on this
particular issue. However, I am quickly running out of time and resources to do so, along with all the other issues
that need to be addressed before a release can be made.
Hence, I favour making what I think is a single innocuous change to throttle back a tiny bit more slowly, where
continual packet loss will always throttle back to absolutely minimum anyway - it may just take slightly longer. On the
surface, a time adjustment appears simpler to me than byte counting, though I haven't thought about that much. There
would be no further changes at this point. Of course, there would be testing but maybe not the extensive and time
expensive load testing that took place during the conference buildup. Although if you have time to help with that it
would be very welcome. If so, I am happy to wait till next week.
I'm also not sure I regard this as a tried and tested algorithm in this context. It certainly is in the TCP world but
there things appear to be rather different - only one segment is going to expire at a time and halve the throttle (not
many UDP acks expiring at once to decimate it). Also, the build up in TCP land looks rather quickly - in our case we
only increase throttle on receipt of an ack. The number of reliable packets sent from server to client is not that high
so the throttle takes a considerable period to build up from low levels.
On 01/12/14 16:42, Mic Bowman wrote:
> one thing that i was concerned about when i put the throttles in place is the relationship between congestion control
> and packet sizes. if you're generating a large number of small, reliable packets that are being dropped, that could
> cause the congestion control to kick in more quickly. that would suggest an adjustment based on bytes sent rather than
> time (though both are probably appropriate).
>
> my biggest concern is that we start fixing by "stabbing in the dark". congestion control is particularly nasty in how it
> interacts which is why i started with a well known & long battle tested algorithm. making random changes might fix one
> problem and introduce a half dozen others.
>
> i'm not in a position to help on the diagnosis until next week if you can wait until then.
>
> --mic
>
>
> On Wed, Nov 26, 2014 at 4:04 PM, Justin Clark-Casey <jjustincc at googlemail.com <mailto:jjustincc at googlemail.com>> wrote:
>
> This was actually happening at quite low loads (< 40 connections over all 4 keynotes). Once adaptive throttles was
> disabled and other unrelated issues fixed the system had no obvious issues coping with higher loads in both testing
> and the conference itself (e.g. the 159 peak keynote avatars in the conference). So I don't think it was a server
> bandwidth issue.
>
> That said, it was somewhat strange behaviour as affected only maybe 10-20% of connections. Once it did affect a
> connection (I saw this happening by logging downward adjustments which one can still do with the console command
> "debug lludp throttles log 1"), the connection would not recover - at some point a bunch of expires would reduce the
> throttle again. Connections seemed to be affected randomly - I experienced the issue myself at one point and I have
> pretty solid fibre.
>
> You're right in that I don't know why this happened or why problematic connections stayed problematic instead of
> slowly recovering. Because of time constraints we had to disable adaptive instead of investigating further. But I
> don't advocate doing this by default at all because, as you say, it's an important mechanism for congestion control.
>
> I do plan further investigation will happen at some point but it's time consuming work and I'd really love to get a
> release out soon-ish. So for the moment I would like to do tune the adapation mechanism tuning as you've mentioned,
> which I believe should probably be done anyway. Because of the nature of the problem, my plan would be not to
> change the adaption divisor but rather to adapt downwards only every 2 seconds or so if packets are expiring rather
> than on every packet expire. I believe this should still achieve the adaption effect without massively penalising
> the connection if there has been a momentary connection issue or similar.
>
> On 26/11/14 02:39, Mic Bowman wrote:
>
> As you mention... cutting the throttle by 50% was modeled on the TCP congestion control approach. It is very
> aggressive
> as a congestion control mechanism and certainly could be tuned.
>
> That being said... do you know why the packets were considered un-acked? If its because the simulator is having
> problems
> (which given your description that it happens under load seems to be the case) then we can probably do something
> more
> intelligent about throttling over all simulator BW. That is... maybe the problem is that the top end of the overall
> simulator bw is the problem, not the per connection throttles.
>
> Manual throttles & adaptive throttles are not exclusive. You can use both. Adaptive manages the top end, but the
> manual
> throttles set an absolute max.
>
> --mic
>
>
> On Tue, Nov 25, 2014 at 5:15 PM, Justin Clark-Casey <jjustincc at googlemail.com <mailto:jjustincc at googlemail.com>
> <mailto:jjustincc at googlemail.__com <mailto:jjustincc at googlemail.com>>> wrote:
>
> Hi Mic (primarily),
>
> Two years ago [1] we had a discussion about the enable_adaptive_throttles setting. Just for background,
> this is a
> setting that adapts the amount of data sent to the viewer depending on whether reliable packets sent from the
> simulator are acked or not. As such, it looks to make sure that a viewer which sets a downstream bandwidth
> higher
> than its network connection can cope with is not permanently hosed with too much data. We enabled it on an
> experimental basis [2].
>
> As you said at the time, this is modelled on the congestion approach used in TCP. I see that for TCP, the
> rate is
> halved on every unacked segment. In OpenSimulator, it's halved on every unacked reliable packet.
>
> However, under fairly modest load conditions in the conference grid, I saw a behaviour where sometimes for a
> connection a sequence of packets would expire for some connections in a very short time period (< 1 sec). This
> would halve the throttle many times, in my observations right down to the absolute minimum. This caused the
> behaviour from the user's point of view to degrade considerably for an extended period of time. The
> throttles takes
> quite a long period to grow again.
>
> I didn't get much further with the diagnostics since a lack of time forced us to switch back to manual
> throttling
> instead (with a 1 mbit per viewer and 400 mbit total on the keynotes). This seemed to work okay in testing
> and in
> the event itself. However, this leaves one vulnerable to the problem adaptive_throttles looks to tackle in the
> first place.
>
> I'm still reading up about this stuff, but it strikes me that halving the throttle on every missed packet
> is much
> harsher than the TCP approach, as with UDP a whole sequence can expire at once rather than a single segment
> that is
> subsequently retried before another segment can be missed.
>
> One idea is to ignore all expiries in a certain period (e.g. next 2 seconds) if an expired packet has
> already caused
> the throttle to be halved. Of course, this is a bit more complicated to do but hopefully not too much so.
> What do
> you think? Any other ideas?
>
> [1] http://opensimulator.org/____pipermail/opensim-dev/2011-____October/023017.html
> <http://opensimulator.org/__pipermail/opensim-dev/2011-__October/023017.html>
> <http://opensimulator.org/__pipermail/opensim-dev/2011-__October/023017.html
> <http://opensimulator.org/pipermail/opensim-dev/2011-October/023017.html>>
> [2] http://opensimulator.org/____pipermail/opensim-dev/2011-____October/023063.html
> <http://opensimulator.org/__pipermail/opensim-dev/2011-__October/023063.html>
> <http://opensimulator.org/__pipermail/opensim-dev/2011-__October/023063.html
> <http://opensimulator.org/pipermail/opensim-dev/2011-October/023063.html>>
>
> Best Regards,
>
> --
> Justin Clark-Casey (justincc)
> OSVW Consulting
> http://justincc.org
> http://twitter.com/justincc
> ___________________________________________________
> Opensim-dev mailing list
> Opensim-dev at opensimulator.org <mailto:Opensim-dev at opensimulator.org> <mailto:Opensim-dev at __opensimulator.org
> <mailto:Opensim-dev at opensimulator.org>>
> http://opensimulator.org/cgi-____bin/mailman/listinfo/opensim-____dev
> <http://opensimulator.org/cgi-__bin/mailman/listinfo/opensim-__dev>
> <http://opensimulator.org/cgi-__bin/mailman/listinfo/opensim-__dev
> <http://opensimulator.org/cgi-bin/mailman/listinfo/opensim-dev>>
>
>
>
>
> _________________________________________________
> Opensim-dev mailing list
> Opensim-dev at opensimulator.org <mailto:Opensim-dev at opensimulator.org>
> http://opensimulator.org/cgi-__bin/mailman/listinfo/opensim-__dev
> <http://opensimulator.org/cgi-bin/mailman/listinfo/opensim-dev>
>
>
>
> --
> Justin Clark-Casey (justincc)
> OSVW Consulting
> http://justincc.org
> http://twitter.com/justincc
> _________________________________________________
> Opensim-dev mailing list
> Opensim-dev at opensimulator.org <mailto:Opensim-dev at opensimulator.org>
> http://opensimulator.org/cgi-__bin/mailman/listinfo/opensim-__dev
> <http://opensimulator.org/cgi-bin/mailman/listinfo/opensim-dev>
>
>
>
>
> _______________________________________________
> Opensim-dev mailing list
> Opensim-dev at opensimulator.org
> http://opensimulator.org/cgi-bin/mailman/listinfo/opensim-dev
>
--
Justin Clark-Casey (justincc)
OSVW Consulting
http://justincc.org
http://twitter.com/justincc
More information about the Opensim-dev
mailing list