Hey guys, i am currently running into issues with the rwin autotuning feature since MS Vista. Here are the key facts which make me get headaches: Microsoft has documented autotuning levels for Vista, default being "normal", others being "restricted" or "experimental" e.g. I was not able to find the same docs for Win7 stack, so i supposed(!!) it was the same which popped up to be not so right in the meantime... First Part of the question: While those levels exist in both Vista and Win7, the advertised WS factor during session initiation is different on both systems. Vista sticks to the ng-stack docs with WS=8 being default and going to e.g. 14 in experimental. The Win7 machines always keep WS=4 and i don't find docs about a change in Win7's TCP Stack.... anyone have an explanation (important: need links quoting the answer like TCP/IP Implementation Docs or CableGuy Article...) Second Part of the question: I have seen Vista machines advertising window sizes of way over 2^16 because of window scaling, even on 100MB links... now in labs I have a physical Vista machine, which ALWAYS maxes its rwin at 2^16, no matter which WS it uses, even with WS=14, rwin is starting at 3 and rises to 4 but never over it. I tried on 100M and Gigabit Links to test if there is a speed derived limit, which exists if you disable autotuning for example, but no difference. -> Does anyone know where this is coming from - did MS Patch s.th. into their ng-stack explaining this behaviour or did I miss some important update on MS's TCP Implementation ? Third Part of the question: Anyone with details about TCP window scaling heuristics, especially what triggers the OS to override the default autotuning-level on a per application base? Any help as always highly appreciated. Thanks in advance!! asked 26 May '11, 02:57 Landi edited 06 Sep '11, 08:49 |
Any ideas if the advertised receive window is related to the round trip time?
Not at all, though I would have expected this because of the bandwidth*delay product, where MS somewhere stated they would address calculation of optimal rwin within Window scaling... obviously doesn't work
push Any TCP heroes around ?