9: Scheduling: Proportional Share

Transcription

1 9: Scheduling: Proortional Share Oerating System: Three Easy Pieces 1

2 Proortional Share Scheduler Fair-share scheduler w Guarantee that each job obtain a certain ercentage of CPU time. w Not otimized for turnaround or resonse time AOS@UC 2

3 Basic Concet Tickets w Reresent the share of a resource that a rocess should receive w The ercent of tickets reresents its share of the system resource in question. Examle w There are two rocesses, A and B. Process A has 75 tickets à receive 75% of the CPU Process B has 25 tickets à receive 25% of the CPU AOS@UC 3

4 Lottery scheduling The scheduler icks a winning ticket. w Load the state of that winning rocess and runs it. Examle w There are 100 tickets Process A has 75 tickets: 0 ~ 74 Process B has 25 tickets: 75 ~ 99 Scheduler s winning tickets: Resulting scheduler: A B A A B A A A A A A B A B A The longer these two jobs comete, The more likely they are to achieve the desired ercentages. AOS@UC 4

5 The beauty of randomness (in scheduling) Deals easily with corner-case situations w Others should have a lot of ifs to revent them Little state required w No need to track the details of each rocess in the ast Really fast w More seed à more seudo-randomness (or HW assistance) AOS@UC 5

6 Ticket Mechanisms Ticket currency w A user allocates tickets among their own jobs in whatever currency they would like. w The system converts the currency into the correct global value. w Examle There are tickets (Global currency) Process A has 100 tickets Process B has 100 tickets User A User B à 500 (A s currency) to A1 à 50 (global currency) à 500 (A s currency) to A2 à 50 (global currency) à 10 (B s currency) to B1 à 100 (global currency) AOS@UC 6

7 Ticket Mechanisms (Cont.) Ticket transfer w A rocess can temorarily hand off its tickets to another rocess. Ticket inflation w A rocess can temorarily raise or lower the number of tickets is owns. w If any one rocess needs more CPU time, it can boost its tickets. AOS@UC 7

8 Imlementation Examle: There are there rocesses, A, B, and C. w Kee the rocesses in a list: head Job:A Tix:100 Job:B Tix:50 Job:C Tix:250 NULL 1 // counter: used to track if we ve found the winner yet 2 int counter = 0; 3 4 // winner: use some call to a random number generator to 5 // get a value, between 0 and the total # of tickets 6 int winner = getrandom(0, totaltickets); 7 8 // current: use this to walk through the list of jobs 9 node_t *current = head; // loo until the sum of ticket values is > the winner 12 while (current) { 13 counter = counter + current->tickets; 14 if (counter > winner) 15 break; // found the winner 16 current = current->next; 17 } 18 // current is the winner: schedule it... AOS@UC 8

9 Imlementation (Cont.) U: unfairness metric w The time the first job comletes divided by the time that the second job comletes. Examle: w There are two jobs, each jobs has runtime 10. First job finishes at time 10 Second job finishes at time 20 w U= "# $# = 0.5 w U will be close to 1 when both jobs finish at nearly the same time. AOS@UC 9

10 Lottery Fairness Study There are two jobs. w Each jobs has the same number of tickets (100). When the job length is not very long, average unfairness can be quite severe. AOS@UC 10

11 Stride Scheduling (deterministic Fair-share scheduler) Stride of each rocess w Defined as (one large number) / (the number of tickets of the rocess) w Examle: one large number = 10,000 Process A has 100 tickets à stride of A is 100 Process B has 50 tickets à stride of B is Process C has 250 tickets à stride of C is 40 A rocess runs, increment a counter(=ass value) for it by its stride. w Pick the rocess to run that has the lowest ass value current = remove_min(queue); // ick client with minimum ass schedule(current); // use resource for quantum current->ass += current->stride; // comute next ass using stride insert(queue, current); // ut back into the queue A seudo code imlementation AOS@UC 11

12 Stride Scheduling Examle Pass(A) (stride=100) Pass(B) (stride=) Pass(C) (stride=40) Who Runs? A B C C C A C C If new job enters with ass value 0, It will monoolize the CPU!: Stride scheduler requires global state (in contrast with lottery) AOS@UC 12

13 Linux Comletely Fair Scheduler (CFS) 5% overall datacenter CPU time wasted in scheduler Focus in fairness w Kee track of the (virtual) runtime of each rocess w At scheduling time, chooses the rocess with the lowest virtual run time w Time slice is variable: according number of ready to run rocesses (from sched_latency (48 ms) for 1 rocess to min_granularity (6 ms) for any number of rocesses Time slice and virtual runtime of the rocess can be affected by weights (niceness) AOS@UC 13

14 Linux Comletely Fair Scheduler (CFS) Minimal scheduler overhead w Data structures should be scalable: No lists w CFS uses a balanced tree (red-black tree) of the ready-to-run rocesses w O(log n) insertions and searches Many more details w Heuristics for multi CPU scheduling (cache affinity) w Cooerative multi-rocess schedule w AOS@UC 14

15 Disclaimer: This lecture slide set is used in AOS course at University of Cantabria. Was initially develoed for Oerating System course in Comuter Science Det. at Hanyang University. This lecture slide set is for OSTEP book written by Remzi and Andrea Araci-Dusseau (at University of Wisconsin) 15