[pjc50]

I worked on a unique HFT system which was capable of starting the response packet before the incoming packet's final byte had arrived. Negative latency. If it mispredicted the future, it would simply scramble the trailer and cause the packet to fail UDP checksum.

It didn't make money in the real world because the quality of decision that could be made at that speed wasn't good enough.

[djoldman]

This was an open secret. You can also send the start of a datagram on every incoming packet and then abort if you don't want to add/modify/delete an order by splitting your datagram into multiple packets. This worked on some exchanges because they ordered add/modify/deletes by when the beginning of the packet was received as opposed to the end.

If you do too much of this you will get angry exchange network people yelling at you as they may consider it spam/ddos. Some exchanges explicitly limit this.

It's also worth considering the possible gain. If you're in the fpga (0-150ns) or cpu(300ns-3us) space, the math can come out differently.

[tijsvd]

This is not unique, it's standard practice. Many exchanges send large UDP packets with the most valuable information at the front. Or the packet is structured such that you can make an informed bet based on size and first sub-message.

Failing the checksum was sort of common as well, but exchanges don't like it.

These days most tricks have to do with avoiding as much serialisation time as possible, e.g. by sending ahead part of the TCP payload and filling in with some innocent order if there was no opportunity.

[twic]

Exchanges have cracked down on this sort of thing a lot this year.

From a letter from the CME to the CFTC dated July 24, 2020:

> On July 26, 2020, an enhancement to the Market Segment Gateway (“MSGW”) will be introduced to further safeguard the CME Globex electronic trading platform (“CME Globex”) infrastructure by introducing a delay of at least three microseconds if the MSGW receives a partial order message as a means of ensuring the stability of the platform. Certain participants intentionally submit partial order messages to reduce latency, and only complete the order message upon the happening of an event or trading signal. Implementation of the enhancement is expected to reduce the frequency of intentionally split order messages as the additional processing time will serve as a deterrent.

The letter is on the web, but it's a PDF; the Google search result has a redirect link to it, DuckDuckGo can't seem to find it, and Firefox on Android won't tell me the URL it downloads things from, so I'm afraid I can't link to it!

CME also made a more general change, where if they decide a participant is sending dodgy messages, they will reroute all their packets to a special gateway for "additional checks", but in practice, to impose a latency penalty. Can't find the documentation on that at all, though.

[eneveu]

Here is the link to the PDF: https://www.cftc.gov/sites/default/files/filings/orgrules/20...

[inopinatus]

From time to time there comes a remark that is the absolute epitome of Hacker News, and I mean that earnestly and without any backhanded rancour, and I thank you for this flawless specimen.

[pjc50]

The original case study: https://web.archive.org/web/20180224141009/http://www.argond...

[KMag]

To get such changes reliably with low latency, I presume you were modifying the DMA buffer, without any syscalls or other mechanisms to synchronize between the CPU and the NIC's processor(s), right? Did you just set up the race condition such that when you got bitten, it resulted in a failed checksum, or was the time window so large that in practice the race condition never bit you?

[pjc50]

It was in a ridiculously expensive 10Gbit ethernet switch with an FPGA built in. Having to write trading algos in Verilog was an extra layer of difficulties.

I don't think it stored the packet at all, just advanced a state machine as each word arrived from the MAC.

[_huayra_]

Wow this is wild! I never thought things would get so latency-sensitive that one would have to overlap request and response!

[guscost]

HFT is so latency-sensitive that it matters how physically close you are on Manhattan island.

It is so latency-sensitive that new “hollow” fiber-optic cables are being installed, because light moves faster in air than in glass.

[vntok]

See also how they slow down the whole network to level the field: https://hackaday.com/2019/02/26/putting-the-brakes-on-high-f...

[KMag]

That's super clever!

[lordnacho]

What was the rest of the infrastructure? Was this at a proper HFT firm with everything else streamlined?

[pjc50]

We were consultants to an HFT firm; supposedly set up correctly but I didn't see that part of the project.