We have worked hard to reduce all forms of pauses, and have GC pauses well in hand even for heaps with 10's of Gigs. Now other forms of pauses dominate.

This talk will discuss some of the issues involved with bringing this stack into Open Source and onto a commodity chip. Business issues: why Open Source? (to break a chicken-and-egg problem: we need OS changes but the OS vendors won't change unless there is a user). Hardware issues: Azul's GC needs Read Barriers and Metadata in Pointers which we used to get from our custom hardware + custom OS stack. How does this work on the X86?. OS issues: We need to (and can) re-map terabytes of virtual memory per second during GC - which we do with a new ABI into the linux kernel. Scheduler issues: The standard linux scheduler gives poor pause-time performance (but great throughput) - and we are attempting a low-pause system and that means soft-real-time. Other causes-for-pauses can come from priority inversion (doesn't happen when you have more CPUs than threads like our custom gear), lock convoying inside the JVM, X86 power management suddenly idling a CPU that is supposed to be advancing a thread to a safepoint, and all sorts of hypervisor interactions.