Linux Scheduler CFS
lottery scheduling
concept
use tickets to represent CPU share, use random, walk through the list until total
count > random number
shortcome:
can’t deliver the exact right properties in short run time
stride scheduling
concept
largen
dividek =
ticketsk
curr = remove_min(queue); // pick client with min pass
schedule(curr); // run for quantum
curr->pass += curr->stride; // update pass using stride
insert(queue, curr); // return curr to queue
shortcome
Linux Scheduler CFS 1
can’t handle mid-in task
CFS
content
choose lowest vruntime task to run
sched_latency
switch_time =
nprocess
to prevent context switching too offen, switch_time has a minimal value
with priority
weight(nice) table
static const int prio_to_weight[40] = {
/* -20 */ 88761, 71755, 56483, 46273, 36291,
/* -15 */ 29154, 23254, 18705, 14949, 11916,
/* -10 */ 9548, 7620, 6100, 4904, 3906,
/* -5 */ 3121, 2501, 1991, 1586, 1277,
/* 0 */ 1024, 820, 655, 526, 423,
/* 5 */ 335, 272, 215, 172, 137,
/* 10 */ 110, 87, 70, 56, 45,
/* 15 */ 36, 29, 23, 18, 15,
};
weightk
time_slicen = ∗ sched_latency
∑n−1
i=0weighti
weight0
vruntimei = vruntimei + ∗ runtimei
weighti
work with i/o
Linux Scheduler CFS 2
job which executes i/o operation will get lower vruntime than others, to prevent
these jobs monopolize the CPU, set job’s vruntine to lowest time in current tree
when waking up
implementation details
use red-black tree to insert job at logarithmic time
Linux Scheduler CFS 3
