Whichever order fulfills the rules of the jurisdiction the exchange resides in, I believe in the US that would mean the order that came in first.
In general orders are serialized first and then processed according to the order that was serialized, it's impossible to reason about an order book with out serializing entries to it.
It's quite easy to come up with good rules that resolve orders without serialization, and certainly without FIFO serialization. There are a number of strategies that could be employed, including "fractional allocation" and "pick randomly". I doubt it really matters exactly what you pick as long as 1) the rules are applied evenly and 2) everyone thinks they are fair.
"My ethernet cable is 5m shorter than your ethernet cable" is does not sound like a reasonable input to the fairness function.
It's easily mitigated by submitting your order 16 nanoseconds sooner.
Imagine UPS complained to the Feds that FedEx had an unfair advantage because they bought faster planes...
And actually it isn't very easy because nothing travels faster than light so you need some ridiculous method to determine with nanosecond accuracy the time that each order was 'created' at, which will invariably lead to people 'creating' orders and then not sending them.
Also the entire reason why LMAX is such a great method is because it IS serialized and IS FIFO doing those two things allows 6 million tps.
It's like seeing a car on the lot announcing I will pay $18,000 for this car and then wondering why the dealer won't sell for less than $18,000.
There's no law that says you have to trade on exchanges with HFT, you can trade on exchanges that don't allow HFT, have fun trying to find a counterparty though. Markets aren't fair they are efficient.
With enough money you can destroy practically any company. Don't like the cut of someone's jib? Buy any debt linked to stock price, short the fuck out of it then call it, leak a story about the outstanding debt and then profit on your short, buy more debt and force bankruptcy as the senior lender.
But "pick randomly" just devolves into pro-rata, which already exists in many products and has its own problems (namely that participants show more than they want). The more volume you show in this scheme, the more volume you get filled in - aka pro-rata.
I'm not sure what fractional allocation is, perhaps that's what I'm calling pro-rata? At any rate, the allocation method that is chosen is hugely important. It will have a large impact on market microstructure.
I find that the discrete mini-auctions are strictly inferior to already existing matching algorithms, as they are really just more complex (read: more gameable by HFT) , yet have the same characteristics of other, more widespread systems.
