> There are probably at least a dozen GC algorithms that provide sub-1ms pause times though
I'd be interested to know more about which ones provide this guarantee. This has been surprisingly difficult to Google for.
I'd be interested to know more about which ones provide this guarantee. This has been surprisingly difficult to Google for.
https://researcher.watson.ibm.com/researcher/view_group_subp...
The Recycler GC is measured around 2ms on old hardware, so it's surely less now because IIRC that pause time is only to copy the stack to a buffer to scan in a separate thread:
Java without the Coffee Breaks: A Nonintrusive Multiprocessor Garbage Collector, https://sites.cs.ucsb.edu/~ckrintz/racelab/gc/papers/bacon-c...
A study of 3 realtime concurrent copying collectors, CHICKEN, CLOVER and STOPLESS, all of which seem to "response times" less than 10 usecs:
A Study of Concurrent Real-Time Garbage Collectors, https://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.353...
Parallel realtime collector, hard to quantify exact pause time because it's configurable and they make a good point that MMU is a better measure:
http://www.aladdin.cs.cmu.edu/papers/pdfs/y2001/rtg.pdf
Then you have various compile-time GC analyses you can apply to your language, like Perceus:
https://www.microsoft.com/en-us/research/uploads/prod/2020/1...
The problem with maximum pause times is that it's not necessarily a feature of all realtime collectors, which try to do work incrementally and so don't necessarily introduce pauses in the same way as traditional tracing GCs. As you can see above, some of them quantify realtime behaviour in various ways so you can trade off memory use, CPU utilization and max latency.
But as mentioned, boundedness is the important property, hard real time systems are not really fast in any way or shape (e.g. for a modern chip to calculate a missile trajectory you really barely need any time, but it not replying in the necessary time window which may be like half a sec, because GC/JIT compiler/OS interrupt is a no-go).