OUTLINE Motivation and Goal Background of NT Eects brought by SMP Methodology Preliminary results 1 Motivation and Goal Hot topic/Mindcraft benchmark Why SMP Approach: benchmarking and analyzing 2 NT: Architecture Object based OS IRQL High Interprocessor Interrupt Clock Device high Device Dispatch/DPC APC Low 3 NT: Process and Thread Process and Thread Scheduling { Priority based multi-queue scheduler { Thread states { AÆnity { Elaborate scheme of selecting proces- sors { Selecting a thread to run a processor 4 NT: Memory Management Address translation Page state: available/Reserved/Committed Variable allocation, local replacement 5 Eects brought by SMP Scheduler(lock contention, processor selection) Memory Management(TLB shoot down) I/O(interrupt handling) File System Network 6 Methodology What is important to SMP? { Science computation { Internet server, esp web server { Database What's the meaning of performance? { General: elapse time and throughput/responsive { Specic to SMP Speedup Cpu utilization (load balance,nature of workload) 7 Methodology we will test: Scheduler Memory system Network system No le system, no I/O system 8 Pitfalls Version selection Things need to be careful { tunable kernel parameters { running demons Benchmarks: workload nature, concurrency Kernel proling 9 Kernel Proling Linux tention: through /proc { /proc/stat { Extention: lockmeter, tlb ush meter, enhanced cpu usage NT side: Kernel Device Driver 10 Preliminary results Speedup on Linux 11