Interview of Cesar Jacquet about technical part
Because of my great interest in Dark and Light, there are many things I would like to know about this famous server which allows us to share fun with a whole community. Together with JoL team, Alderic and I decided to interview Cesar Jacquet NPcube COE.
Below is the interview. If you feel uncomfortable with the technical aspect of the terms used, or need more precisions about some points, feel free to ask your questions on this thread.
< Draltahn > All the equipment is based in Paris, isn’t it?
< César > Yes, we have two rooms near the city of Paris.
< Draltahn > What type of computer do you use for your cluster basis ? Are they specific or commonly used?
< César > Many Bi-Xeon rack is it considered as being commonly used?
< Draltahn > Among all cluster elements, there is an active element (an element which allows information to be transmitted through a network, cables are passive elements) which allows the information to pass?
< César > That’s right, beyond the server, the backbone of a machine room is its network. We have important restraints for breakdown on top of traditional size problem. We have a bottom basket on the switch that keeps the burden, but a cluster like this suffers no error!
The context is really difficult for them because they receive a lot of little trames, which have a high time per machine cost. In our case, the commutation powers prevail over the band switch and require top quality equipments.
< Draltahn > All computers are linked by FO (fibre-optics) at 2Gbit ? (I think that this is the best band switch but I'm not sure)
< César > Internal link between computers of each cluster floor will not be in FO, copper links are more dedicated to that kind of use. FO is not the best for short links in protected environment. But telecom connexions using FO are superfluous.
< Draltahn > So it's 1 Giga bit cupper?
< César > Yes, our specifications and benches on most of links show that 100 Mb are enough, because the CPU servers will be the bottleneck. It's not necessary to overflow the switches with empty links.
< Aldéric > You are right.
In case of technical breakdown, how will you intervene? We had already have the opportunity to face GOA crashes, how do plan to manage this problem ?
< César > Our politics is to make trite all element of the cluster, so we can replace them in hot plug, our server software allows to disconnect an element of the cluster on hot plug mode and we do not have any HDD in our server :)
< Aldéric > How do you manage program storing ?
< César > Configurations are uploaded on unit’s boot, in another mastering and bootstrap cluster. Therefore, we do not have to worry about server installation.
< Aldéric > There is no latency risk? When will the server get data ?
< César > No, it's during the boot process, before the player uses the unit.
< Aldéric > Let’s assume you do not have any HDD data in your server, you must surely have a database for player accounts for example? If this database, or a hard system of this base, breaks down, how will you manage the problem ?
< César > In case of this type of problem, we will switch to another room.
< Aldéric > Given that you will be based far from the technical room which is located in the Reunion Island, I suppose that you have an extension near Paris to manage technical emergencies ?
< César > Yes indeed, we have technicians based in Paris.
< Aldéric > In an other register, some games need daily maintenance, either technical linked to software matters.
How do you plan to manage this frequent intervention ?
< César > We would like to work on a regular basis when providing players with new patches, because of the great impact this efficiency will have on their interest and the game life. But we are still working on identifying the frequency of these interventions.
< Draltahn > Do you have a part that will manage the repartition of the load among all elements, or will a software manage this operation between each computer ?
< César > It deals with equipment on the first levels, but essentially conceptual as far as software is concerned. The load’s balance is in the middle of the evolution of the data traffic inside the cluster.
< Draltahn > Are you linked to several public backbones, or to only one ?
< César > Because it is fundamental, considering the important debit, to have full protection, we have indeed several backbones to optimize the international cover.
< Aldéric > About download platform, client patch cause a lot of lag, have got a solution for this problem?
< César > Our politic is to delegate to our partner, who give us privates mirror sites network.
< Aldéric > Ok, so the download structures in the launcher use all these different mirrors?
< César > Yes you are right
< Aldéric > Do you think the structure can handle the meeting of several hundred people, given that you will need to send to each one of them the information related to hundred of other ones ? For example, a 200 people meeting implies that 200 receive the data of the 199 left. This means a volume that will be 199 times more important than for 200 people, and all this done simultaneously.
< César > The architecture is made for this purpose, we have some techniques dealing with the level of network details according to density and distance... I can't say more! Real test of charge stress would be the final round... but beyond that, game’s options and design are making this much harder…
< Aldéric > According to the Alchemic website, they say that they host you. Is it true ?
< César >No, but we have common partners :-)
We want to thank Cesar for his cooperation and hope that we will have answered most of your questions. :)
27 juillet 2016
6 janvier 2013
21 mars 2011
17 mars 2008
31 octobre 2007