Lag after data centre move

[Flavilandile]

Oh dear, must chime in... Resisting is not possible.

I think we may have found the problem, and funny how the chat was reported to be working even during lag spikes...

In this case, seems that UDP packets are dropped regularly due to PMTUD not working AND there being a MTU bottleneck.

There's lots of reasons why UDP packets can be dropped...

At the top of my head :

At the ingress of a router... because the QOS applied to that ingress say to drop the burst of packets that comes in beyond what has been configured.
At the ingress of a router... because there's just too many packets trying to ingress at the same time, and the QOS say to drop UDP in that case.
At the ingress of a router... because there's just too many packets trying to ingress at the same time and there's no way to recover except by dropping packets.
In the backplane of a router... because the backplane is overloaded by traffic from other cards.
At the egress of a router... All the ingress reasons above.

( rinse/repeat at each router that appears in a traceroute.... and all the ones that do not appear in the traceroute because the packets are in the LDP backbone )

Note that none of them talk about MTU. Because MTU is pointless when you have Terabit Interfaces, any size will go through, even 16Kb jumbo packets.
And actually there's more problems with small packets : 48 bytes to 512 bytes than with 4Kb packets. because small packets comes faster and can overload
the ingress/egress processor while larger packets will give them more time to do their job.

MTU can only be an issue on LANs. But from what I've seen it's probably more some scripts in the game server that don't like the 64bit new systems.

The "correct" way to go about this would probably be to run this thing over DCCP... maybe SCTP or some such might also work better than the current situation... but in the short term (as in less than 5 years ) the best bet would be to just fix the routers so that UDP works again, unreliable though it may be.

Do you know what SCTP is and what it's used for ? Really.... I'm asking that question because I think you're basing your correct way on wrong assumptions regarding that protocol.

You need to create a static link between the two ends (It cannot be dynamic, only the switchover to another link can be dynamic if you have a multihoming configuration ).
Then you have a semi reliable plesiochronous transport tunnel that can be used to carry anything that can be a stream of bytes set up over IP ( it's at the same level as TCP ).
The main use for it is in Telecomunication networks to carry SS7 control messages and to transfer over IP the voice/data/whatever 64K channels of SS7 instead of having to set up PCM links
( T1s in the US, E1s in Europe ). It allows to have a lot more traffic on a cable than PCM can do ( T1 = 1,44Mb; E1 = 2Mb; Ethernet, Optical 1 or 10Gb are common in aggregation, 1Tb is regularly seen in backbones ) [ Sigtran for the old generation equipments and Diameter for the newer ones ]

Edit : oh and doing traceroutes to the gls server is mostly pointless ( except for the few cases where there's a network issue outside of Turbine ), what we need is having the servers answer to the traceroutes ( they have stopped doing that in 2011 ) and their IPs ( it changed with the datacenter move )