Performance of round-robin scheduling: LAST LECTURE ➜ Average waiting time: not optimal ➜ Performance depends heavily on size of time-quantum: Scheduling Algorithms: - too short: overhead for context switch becomes too ➜ Slide 1 FIFO Slide 3 expensive ➜ Shortest job next - too large: degenerates to FCFS policy ➜ Shortest remaining job next - rule of thumb: about 80% of all bursts should be shorter than ➜ Highest Response Rate Next (HRRN) 1 time quantum ➜ no starvation ROUND-ROBIN PRIORITIES quantum=1: ➜ A Each thread is associated with a priority B ➜ Basic mechanism to influence scheduler decision: C D • Scheduler will always choose a thread of higher priority over E one of lower priority 0 5 10 15 20 • Implemented via multiple FCFS ready queues (one per Slide 2 quantum=4: Slide 4 priority) A ➜ Lower-priority may suffer starvation B C • adapt priority based on thread’s age or execution history D E ➜ Priorities can be defined internally or externally 0 5 10 15 20 - internal: e.g., memory requirements, I/O bound vs CPU ➜ Scheduled thread is given a time slice bound ➜ Running thread is preempted upon clock interrupt, running - external: e.g., importance of thread, importance of user thread is returned to ready queue ROUND-ROBIN 1 PRIORITIES 2 Feedback scheduling: RQ0 Release Admit Processor Priority queueing: Queue Runable processes headers RQ1 Release Processor Priority 4 (Highest priority) Slide 5 Slide 7 ¥ ¥ ¥ Priority 3 ¥ ¥ ¥ Priority 2 RQn Release Processor Priority 1 (Lowest priority) Priorities influence access to resources, but do not guarantee a certain fraction of the resource (CPU etc)! Feedback scheduling: 0 5 10 15 20 LOTTERY SCHEDULING A ➜ process gets “lottery tickets” for various resources Feedback B q = 1 C ➜ more lottery tickets imply better access to resource D E Advantages: Slide 6 A Slide 8 Feedback i B ➜ Simple q = 2 C D ➜ Highly responsive E ➜ Allows cooperating processes/threads to implement individual ➜ Penalize jobs that have been running longer scheduling policy (exchange of tickets) i ➜ q =2 : longer time slice for lower priority (reduce starvation) PRIORITIES 3 LOTTERY SCHEDULING 4 Example (taken from Embedded Systems Programming: Four processes a running concurrently ➜ Process A: 15% of CPU time Note: UNIX traditionally uses counter-intuitive priority ➜ Process B: 25% of CPU time representation (higher value = less priority) ➜ Process C: 5% of CPU time Bands: ➜ Process D: 55% of CPU time ➜ Decreasing order of priority Slide 9 How many tickets should each process get to achieve this? Slide 11 • Swapper Number of tickets in proportion to CPU time, e.g., if we have • Block I/O device control • 20 tickets overall File manipulation • Character I/O device control ➜ Process A: 15% of tickets: 3 • User processes ➜ Process B: 25% of tickets: 5 ➜ Process C: 5% of tickets: 1 ➜ Process D: 55% of tickets: 11 TRADITIONAL UNIX SCHEDULING (SVR3, 4.3 BSD) Objectives: Advantages: ➜ support for time sharing ➜ ➜ good response time for interactive users relatively simple, effective ➜ ➜ support for low-priority background jobs works well for single processor systems Disadvantages: Slide 10 Strategy: Slide 12 ➜ ➜ Multilevel feedback using round robin within priorities significant overhead in large systems (recomputing priorities) ➜ ➜ Priorities are recomputed once per second response time not guaranteed ➜ non-preemptive kernel: lower priority process in kernel mode • Base priority divides all processes into fixed bands of priority can delay high-priority process levels • Priority adjustment capped to keep processes within bands ➜ Favours I/O-bound over CPU-bound processes TRADITIONAL UNIX SCHEDULING (SVR3,4.3BSD) 5 NON-PREEMPTIVE VS PREEMPTIVE KERNEL 6 Local memory CPU Complete system NON-PREEMPTIVE VS PREEMPTIVE KERNEL M M M M C C C C C C C C C+ M C+ M C+ M ➜ kernel data structures have to be protected C M C Inter- C M Shared C ➜ connect Internet basically, two ways to solve the problem: memory M C C M C C • Non-preemptive: disable all (most) interrupts while in kernel C C C C C C C C C+ M C+ M C+ M mode, so no other thread can get into kernel mode while in M M M M critial section (a) (b) (c) Slide 13 - Priority inversion possible Slide 15 - Coarse grained (b) Loosely coupled multiprocessor - Works only for uniprocessor • Each processor has its own memory and I/O channels • Preemptive: just lock kernel data structure which is currently • Generally called a distributed memory multiprocessor modified (c) Distributed System - More fine-grained • complete computer systems connected via wide area - Introduces additional overhead, can reduce throughput network • communicate via message passing MULTIPROCESSOR SCHEDULING What kind of systems and applications are there? PARALLELISM Classification of Multiprocessor Systems: Independent parallelism: Local memory CPU Complete system ➜ Separate applications/jobs M M M M ➜ No synchronization C C C C C C C C C+ M C+ M C+ M ➜ Parallelism improves throughput, responsiveness C M C Inter- C M Shared C connect Internet ➜ Slide 14 memory M C C M Slide 16 Parallelism doesn’t affect execution time of (single threaded) C C C C C C programs C C C C C+ M C+ M C+ M M M M M Coarse and very coarse-grained parallelism: (a) (b) (c) ➜ Synchronization among processes is infrequent (a) Tightly coupled multiprocessing ➜ Good for loosely coupled multiprocessors • Processors share main memory, controlled by single • Can be ported to multiprocessor with little change operating system, called symmetric multi-processor (SMP) system MULTIPROCESSOR SCHEDULING 7 PARALLELISM 8 Medium-grained parallelism: ASSIGNMENT OF THREADS TO PROCESSORS ➜ Parallel processing within a single application ➜ • Application runs as multithreaded process Treat processors as a pooled resource and assign threads to processors on demand ➜ Threads usually interact frequently ➜ Good for SMP systems • Permanently assign threads to a processor ➜ Unsuitable for loosely-coupled systems - Dedicate short-term queue for each processor ✔ Slide 17 Slide 19 Low overhead Fine-grained parallelism: ✖ Processor could be idle while another processor has a ➜ Highly parallel applications backlog • e.g., parallel execution of loop iterations • Dynamically assign process to a processor ✖ ➜ Very frequent synchronisation higher overhead ✖ poor locality ➜ Works only well on special hardware ✔ better load balancing • vector computers, symmetric multithreading (SMT) hardware MULTIPROCESSOR SCHEDULING Multiprocessor Scheduling: ASSIGNMENT OF THREADS TO PROCESSORS Which process should be run next and where? Who decides which thread runs on which processor? We discuss: Master/slave architecture: ➜ Tightly coupled multiprocessing ➜ Key kernel functions always run on a particular processor ➜ Very coarse to medium grained parallelism Slide 18 Slide 20 ➜ Master is responsible for scheduling ➜ Homogeneous systems (all processors have same specs, access ➜ Slave sends service request to the master to devices) ✔ simple Design Issues: ✔ one processor has control of all resources, no synchronisation ➜ How to assign processes/threads to the available processors? ✖ Failure of master brings down whole system ➜ Multiprogramming on individual processors? ✖ Master can become a performance bottleneck ➜ Which scheduling strategy ? ➜ Scheduling dependend processes ASSIGNMENT OF THREADS TO PROCESSORS 9 ASSIGNMENT OF THREADS TO PROCESSORS 10 LOAD SHARING: TIME SHARING 0 1 2 3 0 1 2 3 0 1 2 3 4 5 6 7 CPU A 5 6 7 A 5 6 7 CPU 12 8 9 10 11 CPU 4 8 9 10 11 goes idle 8 9 10 11 goes idle 12 13 14 15 12 13 14 15 B 13 14 15 Priority Priority Priority Peer architecture: 7 A B C 7 B C 7 C 6 D E 6 D E 6 D E 5 F 5 F 5 F ➜ Operating system can execute on any processor 4 4 4 3 G H I 3 G H I 3 G H I ➜ 2 J K 2 J K 2 J K Each processor does self-scheduling 1 1 1 ➜ 0 L M N 0 L M N 0 L M N Slide 21 Complicates the operating system Slide 22 (a) (b) (c) - Make sure no two processors schedule the same thread ➜ Load is distributed evenly across the processors - Synchronise access to resources ➜ Use global ready queue ➜ Proper symmetric multiprocessing • Threads are not assigned to a particular processor • Scheduler picks any ready thread (according to scheduling policy) • Actual scheduling policy less important than on uniprocessor ➜ No centralized scheduler required Disadvantages of time sharing: ➜ Central queue needs mutual exclusion • Potential race condition when several CPUs are trying to pick a thread from ready queue • May be a bottleneck blocking processors Slide 23 ➜ Preempted threads are unlikely to resume execution on the same processor • Cache use is less efficient, bad locality ➜ Different threads of same process unlikely to execute in parallel • Potentially high intra-process communication latency ASSIGNMENT OF THREADS TO PROCESSORS 11 LOAD SHARING: TIME SHARING 12 GANG SCHEDULING Combined time and space sharing: ➜ Simultaneous scheduling of threads that make up a single process Thread A running 0 ➜ Useful for applications where performance severely degrades when any part of the application is not running A B A B A B CPU 0 0 0 0 0 0 0 • e.g., often need to synchronise with each other Request 2 Request 1 CPU Slide 24 Reply 1 Reply 2 Slide 26 0 1 2 3 4 5 CPU 1 B1 A1 B1 A1 B1 A1 0 A0 A1 A2 A3 A4 A5 C 1 B0 B1 B2 0 C1 C2 Time 0 100 200 300 400 500 600 2 D0 D1 D2 D3 D4 E0 Time 3 E1 E2 E3 E4 E5 E6 slot 4 A0 A1 A2 A3 A4 A5 5 B0 B1 B2 C0 C1 C2 6 D0 D1 D2 D3 D4 E0 7 E1 E2 E3 E4 E5 E6 SMP SUPPORT IN MODERN GENERAL PURPOSE OS’Sa LOAD SHARING: SPACE SHARING ➜ Solaris 8.0: up to 128 ➜ Linux 2.4: up to 32 Scheduling