Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * IPVS: Shortest Expected Delay 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 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 | |
59c51591 | 19 | * jobs on the ith server and Ui is the fixed service rate (weight) of |
1da177e4 LT |
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 | ||
1da177e4 LT |
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 | } | |
68888d10 | 87 | IP_VS_ERR_RL("SED: no destination available\n"); |
1da177e4 LT |
88 | return NULL; |
89 | ||
90 | /* | |
91 | * Find the destination with the least load. | |
92 | */ | |
93 | nextstage: | |
94 | list_for_each_entry_continue(dest, &svc->destinations, n_list) { | |
95 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) | |
96 | continue; | |
97 | doh = ip_vs_sed_dest_overhead(dest); | |
98 | if (loh * atomic_read(&dest->weight) > | |
99 | doh * atomic_read(&least->weight)) { | |
100 | least = dest; | |
101 | loh = doh; | |
102 | } | |
103 | } | |
104 | ||
b14198f6 JV |
105 | IP_VS_DBG_BUF(6, "SED: server %s:%u " |
106 | "activeconns %d refcnt %d weight %d overhead %d\n", | |
107 | IP_VS_DBG_ADDR(svc->af, &least->addr), ntohs(least->port), | |
108 | atomic_read(&least->activeconns), | |
109 | atomic_read(&least->refcnt), | |
110 | atomic_read(&least->weight), loh); | |
1da177e4 LT |
111 | |
112 | return least; | |
113 | } | |
114 | ||
115 | ||
116 | static struct ip_vs_scheduler ip_vs_sed_scheduler = | |
117 | { | |
118 | .name = "sed", | |
119 | .refcnt = ATOMIC_INIT(0), | |
120 | .module = THIS_MODULE, | |
d149ccc9 | 121 | .n_list = LIST_HEAD_INIT(ip_vs_sed_scheduler.n_list), |
1da177e4 LT |
122 | .schedule = ip_vs_sed_schedule, |
123 | }; | |
124 | ||
125 | ||
126 | static int __init ip_vs_sed_init(void) | |
127 | { | |
1da177e4 LT |
128 | return register_ip_vs_scheduler(&ip_vs_sed_scheduler); |
129 | } | |
130 | ||
131 | static void __exit ip_vs_sed_cleanup(void) | |
132 | { | |
133 | unregister_ip_vs_scheduler(&ip_vs_sed_scheduler); | |
134 | } | |
135 | ||
136 | module_init(ip_vs_sed_init); | |
137 | module_exit(ip_vs_sed_cleanup); | |
138 | MODULE_LICENSE("GPL"); |