Threads and Input/Output in the Synthesis Kernel

H. Massalin, and C. Pu. Proceedings of the 12th ACM Symposium on Operating Systems Principles December 1989, pp. 191-200

The Goals: Make kernel calls fast, lower synchronization overhead, do fine-grain scheduling, adapt to load changes

How:

• Run-time compilation: generate code with constants compiled in for frequent operations. E.g.: each thread only accesses only its own element of a table, so just go for it without locking.
• Some clever optimistic locking. Specialized queue structures for certain access patterns.
• Single address space makes kernel boundary crossings cheap.
• Save partial register set on context switch; generate code to jump right into new thread from old one.
• Interrupts handled by current thread

Some pitfalls: No virtual memory; context switching speed obtained by using RR scheduling (but with variable quanta)