Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * IPVS: Shortest Expected Delay scheduling module |
| 3 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 |
Michael Opdenacker | 59c5159 | 2007-05-09 08:57:56 +0200 | [diff] [blame] | 19 | * jobs on the ith server and Ui is the fixed service rate (weight) of |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 int |
| 45 | ip_vs_sed_init_svc(struct ip_vs_service *svc) |
| 46 | { |
| 47 | return 0; |
| 48 | } |
| 49 | |
| 50 | |
| 51 | static int |
| 52 | ip_vs_sed_done_svc(struct ip_vs_service *svc) |
| 53 | { |
| 54 | return 0; |
| 55 | } |
| 56 | |
| 57 | |
| 58 | static int |
| 59 | ip_vs_sed_update_svc(struct ip_vs_service *svc) |
| 60 | { |
| 61 | return 0; |
| 62 | } |
| 63 | |
| 64 | |
| 65 | static inline unsigned int |
| 66 | ip_vs_sed_dest_overhead(struct ip_vs_dest *dest) |
| 67 | { |
| 68 | /* |
| 69 | * We only use the active connection number in the cost |
| 70 | * calculation here. |
| 71 | */ |
| 72 | return atomic_read(&dest->activeconns) + 1; |
| 73 | } |
| 74 | |
| 75 | |
| 76 | /* |
| 77 | * Weighted Least Connection scheduling |
| 78 | */ |
| 79 | static struct ip_vs_dest * |
| 80 | ip_vs_sed_schedule(struct ip_vs_service *svc, const struct sk_buff *skb) |
| 81 | { |
| 82 | struct ip_vs_dest *dest, *least; |
| 83 | unsigned int loh, doh; |
| 84 | |
| 85 | IP_VS_DBG(6, "ip_vs_sed_schedule(): Scheduling...\n"); |
| 86 | |
| 87 | /* |
| 88 | * We calculate the load of each dest server as follows: |
| 89 | * (server expected overhead) / dest->weight |
| 90 | * |
| 91 | * Remember -- no floats in kernel mode!!! |
| 92 | * The comparison of h1*w2 > h2*w1 is equivalent to that of |
| 93 | * h1/w1 > h2/w2 |
| 94 | * if every weight is larger than zero. |
| 95 | * |
| 96 | * The server with weight=0 is quiesced and will not receive any |
| 97 | * new connections. |
| 98 | */ |
| 99 | |
| 100 | list_for_each_entry(dest, &svc->destinations, n_list) { |
| 101 | if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) && |
| 102 | atomic_read(&dest->weight) > 0) { |
| 103 | least = dest; |
| 104 | loh = ip_vs_sed_dest_overhead(least); |
| 105 | goto nextstage; |
| 106 | } |
| 107 | } |
| 108 | return NULL; |
| 109 | |
| 110 | /* |
| 111 | * Find the destination with the least load. |
| 112 | */ |
| 113 | nextstage: |
| 114 | list_for_each_entry_continue(dest, &svc->destinations, n_list) { |
| 115 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 116 | continue; |
| 117 | doh = ip_vs_sed_dest_overhead(dest); |
| 118 | if (loh * atomic_read(&dest->weight) > |
| 119 | doh * atomic_read(&least->weight)) { |
| 120 | least = dest; |
| 121 | loh = doh; |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | IP_VS_DBG(6, "SED: server %u.%u.%u.%u:%u " |
| 126 | "activeconns %d refcnt %d weight %d overhead %d\n", |
| 127 | NIPQUAD(least->addr), ntohs(least->port), |
| 128 | atomic_read(&least->activeconns), |
| 129 | atomic_read(&least->refcnt), |
| 130 | atomic_read(&least->weight), loh); |
| 131 | |
| 132 | return least; |
| 133 | } |
| 134 | |
| 135 | |
| 136 | static struct ip_vs_scheduler ip_vs_sed_scheduler = |
| 137 | { |
| 138 | .name = "sed", |
| 139 | .refcnt = ATOMIC_INIT(0), |
| 140 | .module = THIS_MODULE, |
Sven Wegener | d149ccc | 2008-08-10 09:18:02 +0000 | [diff] [blame^] | 141 | .n_list = LIST_HEAD_INIT(ip_vs_sed_scheduler.n_list), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 142 | .init_service = ip_vs_sed_init_svc, |
| 143 | .done_service = ip_vs_sed_done_svc, |
| 144 | .update_service = ip_vs_sed_update_svc, |
| 145 | .schedule = ip_vs_sed_schedule, |
| 146 | }; |
| 147 | |
| 148 | |
| 149 | static int __init ip_vs_sed_init(void) |
| 150 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | return register_ip_vs_scheduler(&ip_vs_sed_scheduler); |
| 152 | } |
| 153 | |
| 154 | static void __exit ip_vs_sed_cleanup(void) |
| 155 | { |
| 156 | unregister_ip_vs_scheduler(&ip_vs_sed_scheduler); |
| 157 | } |
| 158 | |
| 159 | module_init(ip_vs_sed_init); |
| 160 | module_exit(ip_vs_sed_cleanup); |
| 161 | MODULE_LICENSE("GPL"); |