Process Run States (Simplified)

Transcription

1 PU Scheduling (S S) Ken Wong Washington University Process Run States (Simplified) asic process run states Parent Forks New Preempt I/O or vent omplete Ready ispatch xit Running Zombie I/O or vent Wait process is blocked Other possible states» Stopped: Not terminated, but not to be scheduled»zombie -Ken Wong, Oct 8 New Jobs -Ken Wong, Oct 8 Short-Term Scheduling RY Queue Preemption or Yield Scheduler Resource Manager locked Jobs PU Request one Long-term scheduler decides which processes should be scheduled by short-term scheduler -Ken Wong, Oct 8 Scheduler Simple Scheduler Tasks»etermine the order in which active processes should contend for the PU(s)»ontext switch between one process and another n Save current process PU state in its Process ontrol lock (P) n Load PU registers from new process P»Run the Idle Process if no runnable processes Scheduler rchitectures» Separate scheduler process (e.g., Unix) n Scheduler kernel process runs when a process blocks, is interrupted (e.g., quantum expires), or is awakened»mbedded scheduler function in each thread n e.g., Windows n Thread enters kernel mode and runs scheduler function to switch context

2 Multiprocessor Scheduling Scheduling is two-dimensional»which process to run?»which PU should run the process? ifficulties»potential contention for scheduling data structures» process holding a spin lock loses the PU and blocks other processes»the cache of a PU that has run a process for a long time often has useful data»the TL of a PU same as above» group of related processes would finish faster if scheduled together n e.g., Parallel make command atched Workload xample Process rrival Service emand +ε +ε +ε 8+ε -Ken Wong, Oct 8 -Ken Wong, Oct 8 -Ken Wong, Oct 8 PU Scheduling Policies Non-Preemptive (can t interrupt process)»ffs (First-ome-First-Served)»SJF (Shortest Job First) or SJN (... Next)»xternal Priority n Static: Priority is assigned once n ynamic: Priority can change during PU usage» F (arliest eadline First) Preemptive (can interrupt process)» Round-Robin n quitably distribute PU time among all processes by giving a time slice (quantum) to each RY process»others: SJF or SJN, Priority, F 8 -Ken Wong, Oct 8 Non-Preemptive Scheduling FFS(FIFO) SJF First-ome- First-Served (smallest arrival time first) Gant hart Shortest Job First (smallest service demand)

3 PU Job Performance Parameters T: Observation period : Number of departures in the interval [,T] : usy period d(i): Service demand of ith arrival t(i): time of the ith departure» job departed job arrived to PU»Interactive jobs: response time s(i): ccumulated service time of ith departure»total time job was in the RUN state (using the PU) w(i): /Queueing time of ith departure»total time job spent in the RY queue 9 -Ken Wong, Oct 8 verage Performance Metrics Notation: verage verage Service verage Throughput (eparture Rate) Utilization -Ken Wong, Oct 8 i n = = x( + ) i= x( i) when there are n jobs t = t(+)/ s = s(+)/ w = w(+)/ w = t s r = /T u = /T FFS Performance of FFS and SJF Service Processes 9 vg.. 8. ontext Switches FFS versus Round-Robin FFS(FIFO) First-ome- First-Served SJF 9... RR(q=) 8 Round- Robin (cycle thru RY queue) -Ken Wong, Oct 8 -Ken Wong, Oct 8

4 FFS versus Round-Robin Round-Robin Scheduling Service FFS RR(q=) Processes ontext vg. Switches N processes will get (/N)th of PU time new process is placed at the end of the RUN/RY queue ffect of context switching» = ontext switch overhead» ach of N processes will get q seconds of PU service and incur seconds of overhead N(q+) seconds to serve N processes once Implementation»Set timer to interrupt every q seconds»r interrupt handler calls scheduler to start next process -Ken Wong, Oct 8 -Ken Wong, Oct 8 -Ken Wong, Oct 8 #Jobs FFS Scheduling 8 Last job finishes at time (=Sum of demands) % utilization during the period [,] vg service time = (++++)/ s = vg turnaround (wait+service) time t = 8.» = (++9++)/ vg wait time = (++++)/ w =. Sum s(i) x t(i) = = Non-Preemptive SJF Scheduling 8 Last job finishes at time (=Sum of demands) % utilization during the period [,] vg service time = (++++)/ s = vg turnaround (wait+service) time t = 8..» = (++9++)/ (++++)/ not generally vg wait time w =. true Sum s(i) x t(i) = (surprising?)» = = Ken Wong, Oct 8 #Jobs

5 lternative Scheduling Policies Job Selection ecision Mode Throughput Response time Overhead Fairness Starvation -Ken Wong, Oct 8 FFS Min arrival time Nonpreemptive -- High if large variance in s(i) Minimum an penalize short jobs No onstant Preemptive Lower for smaller quantum Good for short jobs epends on q Fair No RR Min s(i) SJF Nonpreemptive High for short jobs Good for short jobs an be high Penalizes long jobs Possible 8 -Ken Wong, Oct 8 Shortest Process Next (SPN) xtend batched SJF idea to interactive system»sjf has minimum average turnaround time Interactive Process»Wait for command; xecute ommand; Wait»Treat each command as a job; choose shortest first»which one is the shortest job??? Select jobs with shortest estimated burst time»aging or exponential averaging n stimate based on past behavior»stimate: t (new) = at + (-a) x t, a n t : Previous estimate based on aging formula n t: New measured usage small a? a=?»asy to implement when a = ½ large a? a=? n t (new) = (t + t) >> xponential verage xample ata:, 9,,,,,, xponential verage t (new), a = /»( + )/ =»(9 + )/ =.»(8 +.)/ =.»»( +.9)/ =.98»( +.98)/ =.99»( +.99)/ =.»( +.)/ =.» xponential average converges toward 9 -Ken Wong, Oct 8 Fairness??? One efinition»if there are N users, each user gets /N of the PU»an generalize to giving user i w(i) of the PU where w() w(n) = lgorithm»user i gets K(i) tickets in proportion to w(i) periodically»ach time slice (quantum) is worth Q tokens» user gets its time slice of the PU if K(i) Q n K(i) is reduced by Q every time user i uses a quantum» Service users in round-robin order lgorithm (statistical version)»number the tickets»randomly pick a ticket number to give service -Ken Wong, Oct 8

6 -Ken Wong, Oct 8 Traditional Unix Scheduling System V (Release ),. S Target: Interactive, time-sharing system»good response time for interactive users»long running, background jobs do not starve»multilevel feedback with round robin (q = sec) within each priority queue ase priority values» ivide all processes into fixed bands of priority levels»'nice' values are restricted to prevent movement out of assigned priority band»ands (highest first): Swapper, lock I/O device, File manipulation, haracter I/O device, User process Hard-clock interrupt every msec»kernel collects usage statistics and can preempt process rrivals -Ken Wong, Oct 8 Multilevel Feedback Queue Low Priority Queue Queue... Queue n PU out epartures Lower priority queues have higher quantums S Unix Priority Formulas Priority value of process in time interval i»p(i) = + U'(i-)/ + nice n : ase priority value of process n U'(i): xponential average of PU utilization of process in time interval i n nice: Nice value of process (user-controllable) between - and normally»smallest value is Highest priority; i.e., schedule process with smallest P(i) first xponentially weighted average utilization of process»u'(i) = U(i)/ + U'(i-)/ n U(i): PU utilization of process in time interval i -Ken Wong, Oct 8