EVE Information Portal

"Hey, look, we reinvented load balancing by recursive bi-section!": http://www.netlib.org/utk/lsi/pcwLSI/text/node253.html

Seriously, though, load balancing/graph coloring/graph partitioning has been researched by the high-performance computing community for literally decades, as well as in the data center and cloud computing environment more recently. As a result, there's a huge body of work, both in terms of publications and source code, that you could draw on here instead of having to rediscover the issues yourselves. As another starting point (outside of netlib, linked above, or google scholar), the Zoltan project out of Sandia National Labs has dealt with this issue head-on, and includes implementations of lots of different load balancing algorithms. It even includes a Python interface...

I think I know why they had you rewrite the blog: All the pictures look like stuff my cat throws up. And only 1 out of 8 is reasonably pink! You better go ask Punkturis how it's done.

Other than that: nice blog, interesting sheet.

This may or may not work for larger-scale wars. It's actually much more important that the system intelligently infer from node reinforcement requests I think when it comes to nullsec battles, because that's a much better indication of where the people with the most information expect the most load - and from those points you can infer spillover load (ok, these three points are to be reinforced: it follows from there that the surrounding systems are likely to see heavy traffic as well).

It does not solve the escalation problem, where a battle that was not pre-planned causes its lag to be exponentially increased because the fight and all surrounding systems are on the same node as people try to get in. Yes, many people titan-bridge in or cyno in, but many people will also be taking the gates from nearby systems. Because surprise load in nullsec tends to be strongly correlated with surprise load for every adjacent system, it's not good for a node that expects to have a quiet day to have all of those systems. That ensures that when a suprise asakai breaks out, that only one or two nodes at most are stressed. Were the nodes in nullsec more randomly allocated, you'd tend to see the fight on one node and each surrounding system on a different node, spreading the load.

This is why the idea of clumping together nearby systems on the same node doesn't make sense to me. The choice being made here seems to be avoiding inexplicable lag (the three ratters in solitude going why the **** is my system lagged) by increasing actual lag. I don't really understand why that tradeoff is being made.

Speaking of reducing the number of systems per server... now that your node mapper is making tidy clusters, could you use it to reduce the tidi in systems around declared "need to reinforce" nodes?

What I'm thinking is a temporary bump up to the load fingerprint in say, all systems within three jumps of a system that's been requested to be reinforced. You know the neighboring systems will be heavily hit as well. So the reinforce request would automatically kick up the fingerprint of nearby systems, and the premapper would allocate to them more server power for that one day. Since the capacity would be drawn from basically the entirety of the rest of New Eden, you could quite effectively improve the gameplay in the systems surrounding a reinforced node without causing much trouble at all for anyone else. The worst case scenario would be that some slightly overloaded server would cause just a teeny bit of tidi during peak hours.

Edit: Blah, Weaselior beat me to it...

How long till CCP puts each solarsystem at least on one physical core?

I mean with systems like HP Moonshot / also Intel announcing Knights Landing you could literally do that in a single server rack (with 7800 cpu cores or so).

We thought that we sometimes and on the same node as some nullsec systems. WH alone is rather not possible to generate TiDi (at least untill every single one corp in WH finds connection to other active WH corp and make small skirmish/siege :P)

So I'm going to quote you so you can't edit out the part where you look silly, and then quote the devblog that explains why you look silly.

My brain hurts...

Good work CCP (I think) :P :D

also just saying that the phrase used being 'empire systems' would most likely refer to both high and lowsec systems, as im fairly certain all of the 4 main factions have lowsec areas. if you consider the area u fly in as being non-empire i suggest looking at the logos on the station toolbar next time you're docked up.

It isn't. Solarsystems that are not geographically connected (WH systems) and non-solarsystem services are loadbalanced differently.