net/netfilter/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 #define KMSG_COMPONENT "IPVS"
  39 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  40
  41 #include <linux/module.h>
  42 #include <linux/kernel.h>
  43
  44 #include <net/ip_vs.h>
  45
  46
  47 static inline int
  48 ip_vs_sed_dest_overhead(struct ip_vs_dest *dest)
  49 {
  50         /*
  51          * We only use the active connection number in the cost
  52          * calculation here.
  53          */
  54         return atomic_read(&dest->activeconns) + 1;
  55 }
  56
  57
  58 /*
  59  *      Weighted Least Connection scheduling
  60  */
  61 static struct ip_vs_dest *
  62 ip_vs_sed_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
  63                    struct ip_vs_iphdr *iph)
  64 {
  65         struct ip_vs_dest *dest, *least;
  66         int loh, doh;
  67
  68         IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
  69
  70         /*
  71          * We calculate the load of each dest server as follows:
  72          *      (server expected overhead) / dest->weight
  73          *
  74          * Remember -- no floats in kernel mode!!!
  75          * The comparison of h1*w2 > h2*w1 is equivalent to that of
  76          *                h1/w1 > h2/w2
  77          * if every weight is larger than zero.
  78          *
  79          * The server with weight=0 is quiesced and will not receive any
  80          * new connections.
  81          */
  82
  83         list_for_each_entry_rcu(dest, &svc->destinations, n_list) {
  84                 if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
  85                     atomic_read(&dest->weight) > 0) {
  86                         least = dest;
  87                         loh = ip_vs_sed_dest_overhead(least);
  88                         goto nextstage;
  89                 }
  90         }
  91         ip_vs_scheduler_err(svc, "no destination available");
  92         return NULL;
  93
  94         /*
  95          *    Find the destination with the least load.
  96          */
  97   nextstage:
  98         list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) {
  99                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
 100                         continue;
 101                 doh = ip_vs_sed_dest_overhead(dest);
 102                 if ((__s64)loh * atomic_read(&dest->weight) >
 103                     (__s64)doh * atomic_read(&least->weight)) {
 104                         least = dest;
 105                         loh = doh;
 106                 }
 107         }
 108
 109         IP_VS_DBG_BUF(6, "SED: server %s:%u "
 110                       "activeconns %d refcnt %d weight %d overhead %d\n",
 111                       IP_VS_DBG_ADDR(least->af, &least->addr),
 112                       ntohs(least->port),
 113                       atomic_read(&least->activeconns),
 114                       atomic_read(&least->refcnt),
 115                       atomic_read(&least->weight), loh);
 116
 117         return least;
 118 }
 119
 120
 121 static struct ip_vs_scheduler ip_vs_sed_scheduler =
 122 {
 123         .name =                 "sed",
 124         .refcnt =               ATOMIC_INIT(0),
 125         .module =               THIS_MODULE,
 126         .n_list =               LIST_HEAD_INIT(ip_vs_sed_scheduler.n_list),
 127         .schedule =             ip_vs_sed_schedule,
 128 };
 129
 130
 131 static int __init ip_vs_sed_init(void)
 132 {
 133         return register_ip_vs_scheduler(&ip_vs_sed_scheduler);
 134 }
 135
 136 static void __exit ip_vs_sed_cleanup(void)
 137 {
 138         unregister_ip_vs_scheduler(&ip_vs_sed_scheduler);
 139         synchronize_rcu();
 140 }
 141
 142 module_init(ip_vs_sed_init);
 143 module_exit(ip_vs_sed_cleanup);
 144 MODULE_LICENSE("GPL");