[nponeccop]

Can modern garbage collectors deal with almost full heaps? There is an urban legend that GC only shows good performance with 50% of memory free, which is certainly unacceptable for microcontrollers. Also, most garbage collectors require boxing which prevents tight packing of heap data. Are all these problems already solved somewhere? Jitter is the least important issue as there are a lot of works on realtime collectors.

[obtu]

Unfortunately I've no idea how the various GCs deal with the trade-offs; my point is that (memory bandwidth / memory size) in these embedded things is high enough that garbage-collection is fast, contrary to popular wisdom. I was looking at OCAPIC for an example, which has a very simple stop and copy gc which takes 1.5ms to collect. The trade-offs would have some 2× impact on some metric, but not change the feasibility of the thing.

[nponeccop]

My point is that 2x slowdown is critical for many systems - embedded, gamedev, HPC, systems software. Controlled measurement of slow-down and heap slack caused by GC is beyond capabilities of most embedded devs. Academia provided us with some experimental results saying that in half-empty heaps GC is lightning fast, but there's no (recent) data for 90% full heaps and for very large heaps.