]> bbs.cooldavid.org Git - net-next-2.6.git/blob - net/ipv4/ipvs/ip_vs_sed.c
ipvs: Only call init_service, update_service and done_service for schedulers if defined
[net-next-2.6.git] / net / ipv4 / ipvs / ip_vs_sed.c
1 /*
2  * IPVS:        Shortest Expected Delay scheduling module
3  *
4  * Authors:     Wensong Zhang <wensong@linuxvirtualserver.org>
5  *
6  *              This program is free software; you can redistribute it and/or
7  *              modify it under the terms of the GNU General Public License
8  *              as published by the Free Software Foundation; either version
9  *              2 of the License, or (at your option) any later version.
10  *
11  * Changes:
12  *
13  */
14
15 /*
16  * The SED algorithm attempts to minimize each job's expected delay until
17  * completion. The expected delay that the job will experience is
18  * (Ci + 1) / Ui if sent to the ith server, in which Ci is the number of
19  * jobs on the ith server and Ui is the fixed service rate (weight) of
20  * the ith server. The SED algorithm adopts a greedy policy that each does
21  * what is in its own best interest, i.e. to join the queue which would
22  * minimize its expected delay of completion.
23  *
24  * See the following paper for more information:
25  * A. Weinrib and S. Shenker, Greed is not enough: Adaptive load sharing
26  * in large heterogeneous systems. In Proceedings IEEE INFOCOM'88,
27  * pages 986-994, 1988.
28  *
29  * Thanks must go to Marko Buuri <marko@buuri.name> for talking SED to me.
30  *
31  * The difference between SED and WLC is that SED includes the incoming
32  * job in the cost function (the increment of 1). SED may outperform
33  * WLC, while scheduling big jobs under larger heterogeneous systems
34  * (the server weight varies a lot).
35  *
36  */
37
38 #include <linux/module.h>
39 #include <linux/kernel.h>
40
41 #include <net/ip_vs.h>
42
43
44 static inline unsigned int
45 ip_vs_sed_dest_overhead(struct ip_vs_dest *dest)
46 {
47         /*
48          * We only use the active connection number in the cost
49          * calculation here.
50          */
51         return atomic_read(&dest->activeconns) + 1;
52 }
53
54
55 /*
56  *      Weighted Least Connection scheduling
57  */
58 static struct ip_vs_dest *
59 ip_vs_sed_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
60 {
61         struct ip_vs_dest *dest, *least;
62         unsigned int loh, doh;
63
64         IP_VS_DBG(6, "ip_vs_sed_schedule(): Scheduling...\n");
65
66         /*
67          * We calculate the load of each dest server as follows:
68          *      (server expected overhead) / dest->weight
69          *
70          * Remember -- no floats in kernel mode!!!
71          * The comparison of h1*w2 > h2*w1 is equivalent to that of
72          *                h1/w1 > h2/w2
73          * if every weight is larger than zero.
74          *
75          * The server with weight=0 is quiesced and will not receive any
76          * new connections.
77          */
78
79         list_for_each_entry(dest, &svc->destinations, n_list) {
80                 if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
81                     atomic_read(&dest->weight) > 0) {
82                         least = dest;
83                         loh = ip_vs_sed_dest_overhead(least);
84                         goto nextstage;
85                 }
86         }
87         return NULL;
88
89         /*
90          *    Find the destination with the least load.
91          */
92   nextstage:
93         list_for_each_entry_continue(dest, &svc->destinations, n_list) {
94                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
95                         continue;
96                 doh = ip_vs_sed_dest_overhead(dest);
97                 if (loh * atomic_read(&dest->weight) >
98                     doh * atomic_read(&least->weight)) {
99                         least = dest;
100                         loh = doh;
101                 }
102         }
103
104         IP_VS_DBG(6, "SED: server %u.%u.%u.%u:%u "
105                   "activeconns %d refcnt %d weight %d overhead %d\n",
106                   NIPQUAD(least->addr), ntohs(least->port),
107                   atomic_read(&least->activeconns),
108                   atomic_read(&least->refcnt),
109                   atomic_read(&least->weight), loh);
110
111         return least;
112 }
113
114
115 static struct ip_vs_scheduler ip_vs_sed_scheduler =
116 {
117         .name =                 "sed",
118         .refcnt =               ATOMIC_INIT(0),
119         .module =               THIS_MODULE,
120         .n_list =               LIST_HEAD_INIT(ip_vs_sed_scheduler.n_list),
121         .schedule =             ip_vs_sed_schedule,
122 };
123
124
125 static int __init ip_vs_sed_init(void)
126 {
127         return register_ip_vs_scheduler(&ip_vs_sed_scheduler);
128 }
129
130 static void __exit ip_vs_sed_cleanup(void)
131 {
132         unregister_ip_vs_scheduler(&ip_vs_sed_scheduler);
133 }
134
135 module_init(ip_vs_sed_init);
136 module_exit(ip_vs_sed_cleanup);
137 MODULE_LICENSE("GPL");