1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * IPVS: Shortest Expected Delay scheduling module
5 * Authors: Wensong Zhang <wensong@linuxvirtualserver.org>
11 * The SED algorithm attempts to minimize each job's expected delay until
12 * completion. The expected delay that the job will experience is
13 * (Ci + 1) / Ui if sent to the ith server, in which Ci is the number of
14 * jobs on the ith server and Ui is the fixed service rate (weight) of
15 * the ith server. The SED algorithm adopts a greedy policy that each does
16 * what is in its own best interest, i.e. to join the queue which would
17 * minimize its expected delay of completion.
19 * See the following paper for more information:
20 * A. Weinrib and S. Shenker, Greed is not enough: Adaptive load sharing
21 * in large heterogeneous systems. In Proceedings IEEE INFOCOM'88,
22 * pages 986-994, 1988.
24 * Thanks must go to Marko Buuri <marko@buuri.name> for talking SED to me.
26 * The difference between SED and WLC is that SED includes the incoming
27 * job in the cost function (the increment of 1). SED may outperform
28 * WLC, while scheduling big jobs under larger heterogeneous systems
29 * (the server weight varies a lot).
33 #define KMSG_COMPONENT "IPVS"
34 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
36 #include <linux/module.h>
37 #include <linux/kernel.h>
39 #include <net/ip_vs.h>
43 ip_vs_sed_dest_overhead(struct ip_vs_dest *dest)
46 * We only use the active connection number in the cost
49 return atomic_read(&dest->activeconns) + 1;
54 * Weighted Least Connection scheduling
56 static struct ip_vs_dest *
57 ip_vs_sed_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
58 struct ip_vs_iphdr *iph)
60 struct ip_vs_dest *dest, *least;
63 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
66 * We calculate the load of each dest server as follows:
67 * (server expected overhead) / dest->weight
69 * Remember -- no floats in kernel mode!!!
70 * The comparison of h1*w2 > h2*w1 is equivalent to that of
72 * if every weight is larger than zero.
74 * The server with weight=0 is quiesced and will not receive any
78 list_for_each_entry_rcu(dest, &svc->destinations, n_list) {
79 if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
80 atomic_read(&dest->weight) > 0) {
82 loh = ip_vs_sed_dest_overhead(least);
86 ip_vs_scheduler_err(svc, "no destination available");
90 * Find the destination with the least load.
93 list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) {
94 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
96 doh = ip_vs_sed_dest_overhead(dest);
97 if ((__s64)loh * atomic_read(&dest->weight) >
98 (__s64)doh * atomic_read(&least->weight)) {
104 IP_VS_DBG_BUF(6, "SED: server %s:%u "
105 "activeconns %d refcnt %d weight %d overhead %d\n",
106 IP_VS_DBG_ADDR(least->af, &least->addr),
108 atomic_read(&least->activeconns),
109 refcount_read(&least->refcnt),
110 atomic_read(&least->weight), loh);
116 static struct ip_vs_scheduler ip_vs_sed_scheduler =
119 .refcnt = ATOMIC_INIT(0),
120 .module = THIS_MODULE,
121 .n_list = LIST_HEAD_INIT(ip_vs_sed_scheduler.n_list),
122 .schedule = ip_vs_sed_schedule,
126 static int __init ip_vs_sed_init(void)
128 return register_ip_vs_scheduler(&ip_vs_sed_scheduler);
131 static void __exit ip_vs_sed_cleanup(void)
133 unregister_ip_vs_scheduler(&ip_vs_sed_scheduler);
137 module_init(ip_vs_sed_init);
138 module_exit(ip_vs_sed_cleanup);
139 MODULE_LICENSE("GPL");