[mangokiwi]

> In fairness they obviously cannot benchmark against unreleased versions of ZGC, but equally, I was expecting/hoping to see benchmarks run at allocation rates that ZGC/Shenandoah are actually rated for today. What they're mostly proving here is that you need generational collection to handle high throughputs on allocation heavy benchmarks like lusearch, rather than anything special about LXR specifically.

They evaluate over a diverse range of 17 benchmarks, though only 4 latency benchmarks. Would have been nice to have more latency benchmarks, I agree.

> It mentions LXR will provide "sufficient responsiveness" without concurrent evacuation, but the whole reason ZGC/Shenandoah were developed was because for some use cases G1 level pause times - even though they're usually quite low - were not considered sufficient.

Right. At the cost of a significantly larger heap size though. Memory is cheap nowadays, yes, but large datacenters want to reduce their (electricity/power) running costs, and so if we can reach similar (or _slightly_ worse) latency requirements (and the paper demonstrates that low pause times do not necessarily translate to low latency) while using less memory, I think that in-of-itself is quite appealing. Also do note that, as mentioned in the paper, LXR is has not seen the benefit of years of tuning and tweaks that production collectors have.

I am quite excited by this new GC and hope to see further work done to improve its performance.