2 * IPVS: Never Queue scheduling module
4 * Authors: Wensong Zhang <wensong@linuxvirtualserver.org>
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.
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).
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.
27 * Thanks must go to Marko Buuri <marko@buuri.name> for talking NQ to me.
29 * The difference between NQ and SED is that NQ can improve overall
34 #include <linux/module.h>
35 #include <linux/kernel.h>
37 #include <net/ip_vs.h>
40 static inline unsigned int
41 ip_vs_nq_dest_overhead(struct ip_vs_dest
*dest
)
44 * We only use the active connection number in the cost
47 return atomic_read(&dest
->activeconns
) + 1;
52 * Weighted Least Connection scheduling
54 static struct ip_vs_dest
*
55 ip_vs_nq_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
)
57 struct ip_vs_dest
*dest
, *least
= NULL
;
58 unsigned int loh
= 0, doh
;
60 IP_VS_DBG(6, "ip_vs_nq_schedule(): Scheduling...\n");
63 * We calculate the load of each dest server as follows:
64 * (server expected overhead) / dest->weight
66 * Remember -- no floats in kernel mode!!!
67 * The comparison of h1*w2 > h2*w1 is equivalent to that of
69 * if every weight is larger than zero.
71 * The server with weight=0 is quiesced and will not receive any
75 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
77 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
||
78 !atomic_read(&dest
->weight
))
81 doh
= ip_vs_nq_dest_overhead(dest
);
83 /* return the server directly if it is idle */
84 if (atomic_read(&dest
->activeconns
) == 0) {
91 (loh
* atomic_read(&dest
->weight
) >
92 doh
* atomic_read(&least
->weight
))) {
102 IP_VS_DBG_BUF(6, "NQ: server %s:%u "
103 "activeconns %d refcnt %d weight %d overhead %d\n",
104 IP_VS_DBG_ADDR(svc
->af
, &least
->addr
), ntohs(least
->port
),
105 atomic_read(&least
->activeconns
),
106 atomic_read(&least
->refcnt
),
107 atomic_read(&least
->weight
), loh
);
113 static struct ip_vs_scheduler ip_vs_nq_scheduler
=
116 .refcnt
= ATOMIC_INIT(0),
117 .module
= THIS_MODULE
,
118 .n_list
= LIST_HEAD_INIT(ip_vs_nq_scheduler
.n_list
),
119 #ifdef CONFIG_IP_VS_IPV6
122 .schedule
= ip_vs_nq_schedule
,
126 static int __init
ip_vs_nq_init(void)
128 return register_ip_vs_scheduler(&ip_vs_nq_scheduler
);
131 static void __exit
ip_vs_nq_cleanup(void)
133 unregister_ip_vs_scheduler(&ip_vs_nq_scheduler
);
136 module_init(ip_vs_nq_init
);
137 module_exit(ip_vs_nq_cleanup
);
138 MODULE_LICENSE("GPL");