> On 25 Jul 2022, at 1:27 pm, Y. Richard Yang <[email protected]> wrote: > > Consider a set of resources {Ui} with application-level dependencies (e.g., > U1 -> U2, U1 -> U3, U3 -> U4). I see two transport-oblivious designs: > - Transport oblivious, batch submission: U1, U2, U3 and U4 are submitted to > the (HTTP) transport (buffers) concurrently. Without knowing the > dependencies, the transport may schedule the transfers in the order of U4, > U3, U2 and U1, resulting in application layer latency. > - Transport oblivious, controlled-release submission: app submits U1; and > when U1 is transferred, submits U2/U3 (i.e., those at the roots of the DAG); > after U3 is transported, submit U4. This will avoid priority reversal but a > problem is that if the TCP congestion window is large enough to accommodate > all U1-U4, it helps to submit all all. > > Any previous work/techniques?
To me, the first question here is whether you need to specify this and lock it down, or whether you can allow implementations to find the best patterns for them (perhaps with some recommendations). Cheers, -- Mark Nottingham https://www.mnot.net/ _______________________________________________ alto mailing list [email protected] https://www.ietf.org/mailman/listinfo/alto
