Re: [tsvwg] UDP Options: on forcing the use of UDP CS=0 in connection with FRAG+LITE

[Joe Touch <[e-mail]>]

> On Jul 1, 2019, at 5:42 PM, C. M. Heard <[e-mail]> wrote:
> 
> The complaint now is that (under some implementation strategies, at least)
> FRAG+LITE and CS=0 may have a modest performance advantage ("matters a bit")

That bit can be significant, depending on how maids are laid out.

> compared to the new proposal with a fully compensated checksum. Is the
> purported performance advantage enough to outweigh the fact that CS=0
> suffers a significantly higher drop rate than a fully compensated checksum?

Again, wer also talking about per frag checksums vs. reassembled packet checksum. The latter is safer, due to potential reassembly errors.

The issue is what kind of errors should we catch? What should we be robust to?

IMO,  detecting address or port errors is very unlikely, given how NATS overwrite CSs.
Detecting reasssemblu errrors is very important, esp. at high data rates.
Tolerating implementation errors that are already a problem for others is very low on the list - those paths should be debugged and fixed, especially when tolerating them means either disabling reassembly checks or doing double the work.

IMO, we should design for the long term, not the transient.

But it’d be useful to hear from as many others as possible on these issues to determine where consensus is.

Joe