CUDA assisted everest reinforced node


(Requiem Jofama) #1

Hi!

Just a thought from participating in large scale battles and TiDi.
GPU assisted computing ( GPGPU) for all/some taylor expansion calculations would be very interesting to see, GPU/CPU can use shared memory, unlike running multiple CPU where every CPU need dedicated ram.

Moving some calculations, especially float point such as taylor expansion for sin/cos tracking, orbit etc could offload the cpu quite a bit.

My thought is for every server tick, the cpu can delegate a clearly defined array of calculations to be done by the GPU from ram, by this it can be ensured the calculations will be done before the servertick time is up.

I think CUDA calculations on a tesla card could really boost performance and since it’s defined what calculations it’s supposed to do, you also know how long the number crunching will take.


(Dark Engraver) #2

I think it was mentioned in some update by ccp falcon that they ordered some to experiment with right after issues started coming up recently


(Old Pervert) #3

Late to this thread, but @Dark_Engraver do you have sauce for that? Interested in reading it. Did a quick google, didn’t see anything noteworthy.

CUDA offloading would certainly work, it has been suggested in the past… but whether or not it works with the specific version of Python is another matter.

Something else that will be a godsend on the client side is upcoming 64 bit clients. Thank bob for that. Obviously doesn’t help server issues, but it does mean we can keep our graphics cranked with 6000 in local.


(Whitehound) #4

The issue with the server performance during large battles ins’t one of parallelism, but of serialization. Events need to be processed in a particular order and this order cannot be ignored. So do players expect that their actions are being processed in the order they are given and the results must reflect this order.

EVE runs off a large database and the events not only need to follow an order, but they also need to end up safely in the database and on the disks of the servers. This requires a fair amount of resource locking and waiting for transactions to complete.

The problem isn’t in the mathematical computations, but with the amount of events of thousands of pilots, where each pilot expect to see all these events of all other players occurring in real-time on their clients, even when they don’t have any participation in most of them.

Calculating a few thousand orbits every second isn’t an issue. x86 CPUs have in fact sin/cos functions builtin.


(system) #5

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.