[Devblog] Tranquility Tech IV!

[Porky_pine]

Chuck away the IBM shite

[Tonto_Auri]

@D_Kins_Cider

No, I am not actually. Part of the issue is that although as a player I may not be part of a scene at one particular moment, I potentially can be in a single tick from that particular moment, and thus at least some data on me and my activities needs to be tracked so that I can be correctly inserted into the scene timeline as I arrive.

In fact, no. All the information that needs to be inserted is taken from you, not from any other scene. Technically (sic!) it is possible to warp from one scene in one system to another scene in another system directly. Game master’s magic can do.

If I’m not in the system, the only way I can suddenly appear on scene is to gate in (then the gating process manages bringing me into the scene) or to cyno jump in (and the cyno process manages that much the same way the gating process does).

This is only because game rules say that. But it has nothing to do with technical possibility.

Tracked (and read and re-arranged in chronological order) I completely get.

This is why you do not appear on grid (in scene) SUDDELY. It takes a few split seconds for both your warp bubble and the grid to synchronize event flow.
As an uneducated guess, at the very lest this sequence of events should happen:

1. The scene you leave gets notified of your intention.
2. Your state is prepared for transfer and any effects you cause on the scene are canceled if applicable.
3. The scene you are entering is notified that you are at the doorstep.
4. Your state is sent to the destination scene and you are placed in the queue to appear at the next tick.
5. The source scene is notified that the target node had accepted your movement.
6. The source scene confirm your leave and plan your disappearance on the next tick.
7. The target scene confirm your entrance and start processing your effects.
8. The source scene gets confirmation and finally removes all traces of your presence.

This all happens every time you transfer between scenes. No matter if it is a warp, or gate jump, or admin teleport.

But there’s processing for each action that can absolutely be done in parallel, as long as the results are then serialised when delivered back to the clients.

Again. The universe runs at the 1Hz speed. Everything that takes place within the scene during that second is processed serially to ensure events continuity. For example of breaking this processing, dig a dev blog about ship positioning issues around Revelarion and how they were narrowed down to processing of the incomplete data sets on client. (By nature, client does not have ALL the data server has. Thus, some discrepancy is unavoidable, the trick is HOW do you deal with it.)

Now, whether or not that’s all possible within the constraints of the game engine (the 1 second ticks, etc), that’s a separate story - there’s a lot of coordination required to separate the processing out and bring the results back in), that why I specifically asked if the server code’s optimised for single threaded performance.

If the code cannot be optimised for multithreaded performance (I understand that it may not be able to, and that’s fine), then raw clock speed is still not the only metric that matters - instruction pipeline depth becomes quite interesting for example - there’s no point going from 3.2GHz to 3.6GHz in a core (a 12.5% increase) if your pipeline depth increases by 20% to do so, as you’re actually taking more cycles to get work done in that case.

This I can’t say, but I guess this is more related to hyper-threading issues of some processors, where cores refuse to operate on half the threads under heavy load.

[CCP_Swift]

Nothing to see here

[D_Kins_Cider]

@Tonto_Auri

This I can’t say, but I guess this is more related to hyper-threading issues of some processors, where cores refuse to operate on half the threads under heavy load.

Hyperthreading is “interesting”. Not all pairs of threads can be executed in parallel by a single core. When I work in the virtualisation world and talk about pCPU:vCPU ratios, we always talk about cores, not threads, because although there’s theoretically twice as many threads as cores, the work multiplier in general purpose workload usage is more like 1.3:1 than the 2:1 that people assume it’ll be. When I say “multi-threaded” here, I mean a single thread per core (there’s more than a few environments I’ve worked in where hyper-threading is specifically disabled because it messes with performance quite severely with some types of workloads).

[Rexxar_Santaro]

Last time I installed a server was back in 2016 and that was kind of a video surveillance server binded to the main data SqlServer. The project was interesting. There were 24 high resolution IP cameras with face, person, car and animal AI detection. Therefore, I can say that high LAN and PCIe bandwidth will be in trend for a very long time. Now I try to track the server industry as much as I have time to.

@CCP_Explorer

7 years actually, that’s the extended warranty.

Is this a big issue for the company? Are there some specific license options for that? Based on Tranquility Tech IV specs, it’s obvious that you try to fit into the licensing mode

“… minimum of 8 licenses per physical processor and a minimum of 16 licenses per server …”.

IDK. That will be nice to have different servers for different nodes. It’s obvious that for Hi-sec (except main trade hub constellations), Low-sec, Pochven, Abyss and WH nodes don’t need very high performance servers. I saw only Null-sec systems (SOL nodes) with high CPU and connection load during a 5 year period. At the same time there is a big investment dilemma for CCP by choosing between performance and economical advisability. Nothing can happen in a random Null-sec system during a long period of time and suddenly, in 60 seconds, like 4000 players will jump there and start shooting.

@CCP_Explorer

Well, there is more to this. The Xeon Platinum processor are very expensive and some of the Gold ones are power-hungry and/or don’t fit our setup. Some liberal colour-coding:

High power CPUs aren’t an issue by handling a few servers only. Moreover, those CPUs have Turbo Boost and not always will run on full load. In addition to Turbo Boost feature I see more and more server CPUs with more Efficient-cores and I want to start a hashtag for #iwannacry-on-$15000-servers.

IDK. Maybe you already know about this, but here’s the tests for generic performance for the new 4th Gen Xeon. Maybe it will be useful for somebody. It’s not only on paper. I see special labs are testing this during the last 7 months. Those CPUs are marked as “confidential”. Interestingly, only Micron has server DDR5.

PHPBench 0.8.1

Summary

Intel's Sapphire Rapids Finally Here! Feat. Supermicro SYS-221BT-DNTR

@CCP_Explorer

It is also possible that we will move to AMD machines next time since IBM/Lenovo will not be making FLEX blades anymore (but purchasing FLEX blades was a requirement for this upgrade due to existing infrastructure) and the AMD EPYC Zen 4 processors are an interesting option.

Experts say that the future is for node-servers. They are more power efficient and compact than current blade-servers. Maybe yes, maybe not… Take a look at for options

TYAN® Computer - HPC, GPU, AI, Cloud Servers

Supermicro Data Center Server, Blade, Data Storage, AI System

[MalcomReynolds_Serenity]

@Rexxar_Santaro

Is this a big issue for the company? Are there some specific license options for that? Based on Tranquility Tech IV specs, it’s obvious that you try to fit into the licensing mode

Was that a serious question? Of course running unsupported hardware is a problem.

@Rexxar_Santaro

High power CPUs aren’t an issue by handling a few servers only. Moreover, those CPUs have Turbo Boost and not always will run on full load. In addition to Turbo Boost feature I see more and more server CPUs with more Efficient-cores and I want to start a hashtag for #iwannacry-on-$15000-servers.

The platinum line of Xeon processors aren’t just higher powered, they are for different applications. Those are designed to go in 8 socket systems. Eve online is not an HPC workload. That particular processor is only nominally faster than the 6334 but radically more expensive. Also as they said elsewhere, it was necessary to disable turbo boost which is not an uncommon requirement for high concurrency applications, especially applications with extreme ranges of operating norms.

@Rexxar_Santaro

Experts say that the future is for node-servers. They are more power efficient and compact than current blade-servers. Maybe yes, maybe not… Take a look at for options

Experts say a lot of things. the node model does not scale down cost effectively. Even in a relatively large server footprint networks of hundreds or thousands of servers, small islands of workloads emerge with different performance requirements and licensing quickly becomes an absolute nightmare and totally abstract. The various hyperconverged platforms I like to call ‘magic boxes’ various vendors provide somewhat solve this, but it is not scalable and not well suited to high disk i/o, high disk capacity applications.

[Apo_Lamperouge]

but can it run Crysis?

[Apo_Lamperouge]

yeah get computer nerds in a spaceship game talking about hardware and watch the chaos