[jcul]

Why not just make all the abandon transactions into fake discarded transactions, discard them at the send later. E.g. by poisoning the frame checksum or setting something invalid on them, so they get discarded.

Seems you'd be doing this anyway with the dummy transactions.

Then you have no branch, though may want to add dummy transactions anyway to keep the code in cache.

[andrepd]

This is literally what is says in TFA lol

[jcul]

I don't believe that's true.

The article suggests flooding the system with dummy "should send" transactions so that they become the majority.

Quote:

> One such solution that I know of is one that Carl Cook talked about during his CppCon 17 talk2: we can fill our system with mocked transaction data for which should_send(t) returns true. We have to do this enough times such that the mocked transaction data becomes the vast majority over the real data, so the branch predictor will be primed to assume that the send() path will be executed practically on every resolve() call. Overall, this may actually be a better approach than hard-coding prediction rules, because those rules wouldn’t really guarantee us that the whole send() path would get executed (the assumption may just lead to partial execution, until the should_send(t) condition is actually evaluated and the pipeline is flushed; so at the end we may still have important stuff not placed in the instruction/data cache).

What I am suggesting is to remove the branch entirely, and instead poison the "should abandon" transactions so they get dropped or null routed on the NIC. This is the kind of thing that low latency cut through L1 switches do.

Thereby removing the CPU branch predictor from the equation entirely.