> This is a fundamental principle of garbage collection. You can either have low latency or high throughput. You can't get both.
> All optimizations that improve latency come at a cost
Huh? That's a very strange, absolutist statement. There are many cases where you have the opportunity to trade off one of latency and throughput for the other, yes. But there are also many cases where an optimization can improve both.
Optimization is probably the wrong word, fundamental algorithm might be a better one. For example, I certainly could see some SIMD optimization improving latency and throughput, that's not what I'm talking about. I'm talking about the overarching GC algorithm.
AFAIK, there's no algorithm that's both good at throughput and latency without making assumptions around how memory is used.
> All optimizations that improve latency come at a cost
Huh? That's a very strange, absolutist statement. There are many cases where you have the opportunity to trade off one of latency and throughput for the other, yes. But there are also many cases where an optimization can improve both.