1/* 2 * IPVS: Never Queue scheduling module 3 * 4 * Version: $Id: ip_vs_nq.c,v 1.1.1.1 2007/08/03 18:53:51 Exp $ 5 * 6 * Authors: Wensong Zhang <wensong@linuxvirtualserver.org> 7 * 8 * This program is free software; you can redistribute it and/or 9 * modify it under the terms of the GNU General Public License 10 * as published by the Free Software Foundation; either version 11 * 2 of the License, or (at your option) any later version. 12 * 13 * Changes: 14 * 15 */ 16 17/* 18 * The NQ algorithm adopts a two-speed model. When there is an idle server 19 * available, the job will be sent to the idle server, instead of waiting 20 * for a fast one. When there is no idle server available, the job will be 21 * sent to the server that minimize its expected delay (The Shortest 22 * Expected Delay scheduling algorithm). 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 NQ to me. 30 * 31 * The difference between NQ and SED is that NQ can improve overall 32 * system utilization. 33 * 34 */ 35 36#include <linux/module.h> 37#include <linux/kernel.h> 38 39#include <net/ip_vs.h> 40 41 42static int 43ip_vs_nq_init_svc(struct ip_vs_service *svc) 44{ 45 return 0; 46} 47 48 49static int 50ip_vs_nq_done_svc(struct ip_vs_service *svc) 51{ 52 return 0; 53} 54 55 56static int 57ip_vs_nq_update_svc(struct ip_vs_service *svc) 58{ 59 return 0; 60} 61 62 63static inline unsigned int 64ip_vs_nq_dest_overhead(struct ip_vs_dest *dest) 65{ 66 /* 67 * We only use the active connection number in the cost 68 * calculation here. 69 */ 70 return atomic_read(&dest->activeconns) + 1; 71} 72 73 74/* 75 * Weighted Least Connection scheduling 76 */ 77static struct ip_vs_dest * 78ip_vs_nq_schedule(struct ip_vs_service *svc, const struct sk_buff *skb) 79{ 80 struct ip_vs_dest *dest, *least = NULL; 81 unsigned int loh = 0, doh; 82 83 IP_VS_DBG(6, "ip_vs_nq_schedule(): Scheduling...\n"); 84 85 /* 86 * We calculate the load of each dest server as follows: 87 * (server expected overhead) / dest->weight 88 * 89 * Remember -- no floats in kernel mode!!! 90 * The comparison of h1*w2 > h2*w1 is equivalent to that of 91 * h1/w1 > h2/w2 92 * if every weight is larger than zero. 93 * 94 * The server with weight=0 is quiesced and will not receive any 95 * new connections. 96 */ 97 98 list_for_each_entry(dest, &svc->destinations, n_list) { 99 100 if (dest->flags & IP_VS_DEST_F_OVERLOAD || 101 !atomic_read(&dest->weight)) 102 continue; 103 104 doh = ip_vs_nq_dest_overhead(dest); 105 106 /* return the server directly if it is idle */ 107 if (atomic_read(&dest->activeconns) == 0) { 108 least = dest; 109 loh = doh; 110 goto out; 111 } 112 113 if (!least || 114 (loh * atomic_read(&dest->weight) > 115 doh * atomic_read(&least->weight))) { 116 least = dest; 117 loh = doh; 118 } 119 } 120 121 if (!least) 122 return NULL; 123 124 out: 125 IP_VS_DBG(6, "NQ: 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 136static struct ip_vs_scheduler ip_vs_nq_scheduler = 137{ 138 .name = "nq", 139 .refcnt = ATOMIC_INIT(0), 140 .module = THIS_MODULE, 141 .init_service = ip_vs_nq_init_svc, 142 .done_service = ip_vs_nq_done_svc, 143 .update_service = ip_vs_nq_update_svc, 144 .schedule = ip_vs_nq_schedule, 145}; 146 147 148static int __init ip_vs_nq_init(void) 149{ 150 INIT_LIST_HEAD(&ip_vs_nq_scheduler.n_list); 151 return register_ip_vs_scheduler(&ip_vs_nq_scheduler); 152} 153 154static void __exit ip_vs_nq_cleanup(void) 155{ 156 unregister_ip_vs_scheduler(&ip_vs_nq_scheduler); 157} 158 159module_init(ip_vs_nq_init); 160module_exit(ip_vs_nq_cleanup); 161MODULE_LICENSE("GPL"); 162