Digging around a bit, this paper seems to be it. One of the authors was also previously involved in a real-time Java VM I hadn’t heard of. They seem to mean hard real-time, with the experiments in the paper being run on RTEMS, a specialized real-time OS, so this is probably not intended for regular applications.
One of the (to me) clever tricks is having a hard max-size allocation chunk. It’s obviously a costly brute-force sort of decision: large objects are represented as linked lists, and arrays as 32-ary trees(!). But it means never having to have to compact to clear out a large contiguous area for a big allocation.
What? Real-time GC? This is big news for SML.
Yeah, this seems understated/underexplained.
I assume this is soft real-time a la the Azul JVM? Does anyone know where the paper/background on this is? Repo didn’t link.
Digging around a bit, this paper seems to be it. One of the authors was also previously involved in a real-time Java VM I hadn’t heard of. They seem to mean hard real-time, with the experiments in the paper being run on RTEMS, a specialized real-time OS, so this is probably not intended for regular applications.
One of the (to me) clever tricks is having a hard max-size allocation chunk. It’s obviously a costly brute-force sort of decision: large objects are represented as linked lists, and arrays as 32-ary trees(!). But it means never having to have to compact to clear out a large contiguous area for a big allocation.