Matthew Toseland wrote: > Oskar tells me that the following will work a lot better than our > current strategy for storing data, according to his simulations:
Well, I said other strategies than the current worked a lot better _in_ my simulations. You can take that as you wished. > We have a separate cache and store. Both are LRU. The cache stores > everything which passes through the node (possibly excluding locally > originated traffic on more paranoid nodes). The store stores ONLY data > from inserts, and it only stores it if the HTL was reset on that node > because it was a new best location for the request. (Is this right > Oskar?). The store would probably be larger than the cache. This isn't exactly what I simulated (though it is similar with the addition below). I got the best results in simulation from: During both inserts and requests, the query continues until it terminates by having gone H steps without finding a new closest node. After the query is finished, and if it was an insert or the data was found, the data is copied to the N nodes found that were closest to the key value. I never simulated a secondary store, I simply suggested that it could be used to help make popular data available without overloading its closest nodes. I don't know if it is a good method: I think I share (what I perceive as) Ian's gut instinct that it would be nice with a single, adaptive, data storage system, but I don't know how to do it. If anybody does, tell me and I'll try it. I used a large H (20) and a small N (3), and this is in a very small (2000 node) highly connected (~30 neighbors average) network. > Most DHTs include a two level store, as oskar pointed out. The best > results on a static network come from storing it only on the single node > closest to the target; storing it on the 3 nodes closest to the target > seemed to work well on a dynamic network. But that's hard to implement, > hence the suggested solution of storing the data on the "peaks". I have simulations on active networks with node churn as well. // oskar
