You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Text: Wulf2k: "I feel the "intention" of the mp, was for everybody to be connected to everybody. If a darkmoon tries to invade and there's a single player online with sin, they should eventually connect.
Matching based on player preference is just a step towards pretending that it works as it should. Keep in mind, all mp requests are sent to your nodes,and then your nodes pass along your request to their nodes.
This is just such a shitty system hat by the time somebody answers, the request has timed out."
So... I'm sure you've already thought about this. "by the time somebody answers, the request has timed out." That is, a request that runs along the collective graph times out too often for the system to work well.
I think DSCM may have the opportunity to patch the request passing routines. Maybe we could find out how the requests measure their own lifetime and patch that in some way, so that they live long enough to traverse a whole bunch of nodes and try to get a positive response.
Speculatively, if the request was seconds based, we could lengthen the lifetime every time the request hit a new node. Or, if it was nodes hit, we could artificially increase it by how many other nodes connected to the current node that the request is on.
Then, the objective of the system would shift toward "Get those responses out and make sure they try hard to come back positive." Then, the client has to deal with the incoming requests, but in return for increasing the graph space traversed...
Well, this is all speculation anyway. I'm not sure where else to put it or how to flag it as low-priority.
The text was updated successfully, but these errors were encountered:
Keep in mind, all mp requests are sent to your nodes,and then your nodes pass along your request to their nodes.
Did you seed that from monitoring the process, or why do you believe that @Wulf2k. I haven't got a mp interaction with a node I wasn't directly connected to yet. My understanding is that you can only interact with your neighboring nodes.
Ancient knowledge from when the game was popular, unconfirmed through looking at the code.
When trying to connect to friends it was incredibly common to spend 10+ minutes trying to summon them off of their sign and fail. You'd see their sign where they'd placed it 5 minutes ago, not where it was currently.
People had entire rituals of placing a sign, lifting it, waiting a minute, placing another, and so on thinking that it would help.
Then, once they had their first successful summon, everything was fine for the rest of the session.
Even recently there have been a few posts about "summoning failed". Asking people confirms that the signs they're trying to summon don't appear as Active Nodes.
Wulf2k, one of your posts made me think of something. https://www.reddit.com/r/darksouls/comments/4invy1/darkmoon_edition_dscm_dark_souls_connectivity_mod/d3050wy
Text: Wulf2k: "I feel the "intention" of the mp, was for everybody to be connected to everybody. If a darkmoon tries to invade and there's a single player online with sin, they should eventually connect.
Matching based on player preference is just a step towards pretending that it works as it should. Keep in mind, all mp requests are sent to your nodes,and then your nodes pass along your request to their nodes.
This is just such a shitty system hat by the time somebody answers, the request has timed out."
So... I'm sure you've already thought about this. "by the time somebody answers, the request has timed out." That is, a request that runs along the collective graph times out too often for the system to work well.
I think DSCM may have the opportunity to patch the request passing routines. Maybe we could find out how the requests measure their own lifetime and patch that in some way, so that they live long enough to traverse a whole bunch of nodes and try to get a positive response.
Speculatively, if the request was seconds based, we could lengthen the lifetime every time the request hit a new node. Or, if it was nodes hit, we could artificially increase it by how many other nodes connected to the current node that the request is on.
Then, the objective of the system would shift toward "Get those responses out and make sure they try hard to come back positive." Then, the client has to deal with the incoming requests, but in return for increasing the graph space traversed...
Well, this is all speculation anyway. I'm not sure where else to put it or how to flag it as low-priority.
The text was updated successfully, but these errors were encountered: