Paolo ArdoinoVerified account

Fun fact: lots of exchanges claim their matching engine can process huge numbers of orders per second. They're basically all lying.

1/2 The lie is more subtle. You can easily create a matching engine that have that perf. 2 red-black tries + binary UDP I/O. The real lie is that most of exchanges can't handle writing, reconciling and distributing data at that speed: internal distribution, DB I/O, user delivery

2/2 Using a cloud provider and not a dedicated infrastructure affects the ability to reach higher speeds as well. AWS disks throughput has bursts, latency between hosts is not under your control, ... Owned racks, while more painful to manage, are better to get that kind of speed

Is this really a problem with nvme discs though? Their IOPS is batshit insane now. Even RAMdiscs, despite being volatile, can be realtime-streamed to a non-volatile mirror backup.

AWS conf is opaque from what I experienced. Unless you know details you can't go HF. We use raid10 (8 disk) setup on all servers. And back to back connectivity is key for efficient replica. If you get a server with 70+ cores + 8 SSDs on AWS you're paying them new every 5 months.

I'm not really coming at it from a cost standpoint, more a theoretical "what is possible with modern hardware". Based on what I've seen, AWS costs are insane, yeah. Is the peering\colo\replication\whatever the term is ever worth it? They've gotta be massive for a reason.