tailslayer: Library for reducing tail latency in RAM reads
39 points by indigo
39 points by indigo
There's a great video going through the development of this: https://youtu.be/KKbgulTp3FE
Whenever I watch one of her videos I'm amazed she's not doing this full time.
But I need to watch this one again. I have not yet understood how these faster reads can have an effect that is not doubled, since there can be no communication between to simultaneous reads without destroying the advantage the removed latency provides.
Near the end of the video, she mentions that you could program smart NICs to drop the late result. Of course, you need a smart NIC to do that but if you're likely to benefit from the tail-latency reduction, you likely have one already.
Ah, that's where she put that logic, thanks.
I wonder how the receiving end of a HFT looks like to be able to react to these requests in a timely manner. But I guess ordering received trades by arrival time is still easier to do than winning the sending race.