SNS - high OS overhead
Price for many network connections
- O(n) select loop (hidden by thread scheduler)
- high network stack overhead (AM, xKernel, …)
- 10 ?sec in-kernel AM trap, 100 ?sec cross-kernel AM read, 3 msec TCP cross-kernel read()
Thread subsystem overhead
- thread context switch 20-50x procedure call
- RISC: large register sets assume procedure calls more frequent than context switches
- TranSend: 25-50 kB per-thread state