Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * IPVS: Never Queue scheduling module | |
3 | * | |
1da177e4 LT |
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 NQ algorithm adopts a two-speed model. When there is an idle server | |
17 | * available, the job will be sent to the idle server, instead of waiting | |
18 | * for a fast one. When there is no idle server available, the job will be | |
19 | * sent to the server that minimize its expected delay (The Shortest | |
20 | * Expected Delay scheduling algorithm). | |
21 | * | |
22 | * See the following paper for more information: | |
23 | * A. Weinrib and S. Shenker, Greed is not enough: Adaptive load sharing | |
24 | * in large heterogeneous systems. In Proceedings IEEE INFOCOM'88, | |
25 | * pages 986-994, 1988. | |
26 | * | |
27 | * Thanks must go to Marko Buuri <marko@buuri.name> for talking NQ to me. | |
28 | * | |
29 | * The difference between NQ and SED is that NQ can improve overall | |
30 | * system utilization. | |
31 | * | |
32 | */ | |
33 | ||
34 | #include <linux/module.h> | |
35 | #include <linux/kernel.h> | |
36 | ||
37 | #include <net/ip_vs.h> | |
38 | ||
39 | ||
40 | static int | |
41 | ip_vs_nq_init_svc(struct ip_vs_service *svc) | |
42 | { | |
43 | return 0; | |
44 | } | |
45 | ||
46 | ||
47 | static int | |
48 | ip_vs_nq_done_svc(struct ip_vs_service *svc) | |
49 | { | |
50 | return 0; | |
51 | } | |
52 | ||
53 | ||
54 | static int | |
55 | ip_vs_nq_update_svc(struct ip_vs_service *svc) | |
56 | { | |
57 | return 0; | |
58 | } | |
59 | ||
60 | ||
61 | static inline unsigned int | |
62 | ip_vs_nq_dest_overhead(struct ip_vs_dest *dest) | |
63 | { | |
64 | /* | |
65 | * We only use the active connection number in the cost | |
66 | * calculation here. | |
67 | */ | |
68 | return atomic_read(&dest->activeconns) + 1; | |
69 | } | |
70 | ||
71 | ||
72 | /* | |
73 | * Weighted Least Connection scheduling | |
74 | */ | |
75 | static struct ip_vs_dest * | |
76 | ip_vs_nq_schedule(struct ip_vs_service *svc, const struct sk_buff *skb) | |
77 | { | |
78 | struct ip_vs_dest *dest, *least = NULL; | |
79 | unsigned int loh = 0, doh; | |
80 | ||
81 | IP_VS_DBG(6, "ip_vs_nq_schedule(): Scheduling...\n"); | |
82 | ||
83 | /* | |
84 | * We calculate the load of each dest server as follows: | |
85 | * (server expected overhead) / dest->weight | |
86 | * | |
87 | * Remember -- no floats in kernel mode!!! | |
88 | * The comparison of h1*w2 > h2*w1 is equivalent to that of | |
89 | * h1/w1 > h2/w2 | |
90 | * if every weight is larger than zero. | |
91 | * | |
92 | * The server with weight=0 is quiesced and will not receive any | |
93 | * new connections. | |
94 | */ | |
95 | ||
96 | list_for_each_entry(dest, &svc->destinations, n_list) { | |
97 | ||
98 | if (dest->flags & IP_VS_DEST_F_OVERLOAD || | |
99 | !atomic_read(&dest->weight)) | |
100 | continue; | |
101 | ||
102 | doh = ip_vs_nq_dest_overhead(dest); | |
103 | ||
104 | /* return the server directly if it is idle */ | |
105 | if (atomic_read(&dest->activeconns) == 0) { | |
106 | least = dest; | |
107 | loh = doh; | |
108 | goto out; | |
109 | } | |
110 | ||
111 | if (!least || | |
112 | (loh * atomic_read(&dest->weight) > | |
113 | doh * atomic_read(&least->weight))) { | |
114 | least = dest; | |
115 | loh = doh; | |
116 | } | |
117 | } | |
118 | ||
119 | if (!least) | |
120 | return NULL; | |
121 | ||
122 | out: | |
123 | IP_VS_DBG(6, "NQ: server %u.%u.%u.%u:%u " | |
124 | "activeconns %d refcnt %d weight %d overhead %d\n", | |
125 | NIPQUAD(least->addr), ntohs(least->port), | |
126 | atomic_read(&least->activeconns), | |
127 | atomic_read(&least->refcnt), | |
128 | atomic_read(&least->weight), loh); | |
129 | ||
130 | return least; | |
131 | } | |
132 | ||
133 | ||
134 | static struct ip_vs_scheduler ip_vs_nq_scheduler = | |
135 | { | |
136 | .name = "nq", | |
137 | .refcnt = ATOMIC_INIT(0), | |
138 | .module = THIS_MODULE, | |
139 | .init_service = ip_vs_nq_init_svc, | |
140 | .done_service = ip_vs_nq_done_svc, | |
141 | .update_service = ip_vs_nq_update_svc, | |
142 | .schedule = ip_vs_nq_schedule, | |
143 | }; | |
144 | ||
145 | ||
146 | static int __init ip_vs_nq_init(void) | |
147 | { | |
148 | INIT_LIST_HEAD(&ip_vs_nq_scheduler.n_list); | |
149 | return register_ip_vs_scheduler(&ip_vs_nq_scheduler); | |
150 | } | |
151 | ||
152 | static void __exit ip_vs_nq_cleanup(void) | |
153 | { | |
154 | unregister_ip_vs_scheduler(&ip_vs_nq_scheduler); | |
155 | } | |
156 | ||
157 | module_init(ip_vs_nq_init); | |
158 | module_exit(ip_vs_nq_cleanup); | |
159 | MODULE_LICENSE("GPL"); |