Message ID | 200910251954.49700.opurdila@ixiacom.com |
---|---|
State | RFC, archived |
Delegated to: | David Miller |
Headers | show |
Very interesting patch, because i have PPPoE and sysctl locking issue my issue N1(accorting to perf and oprofile) on massive pppoe login and during operation. Probably i will try to apply it on one of loaded (but redundant, in case it will crash) pppoe. On Sunday 25 October 2009 19:54:49 Octavian Purdila wrote: > RFC patches are attached. > > Another possible approach: add an interface flag and use it to decide > whether we want per interface sysctl entries or not. > > Benchmarks for creating 1000 interface (with the ndst module previously > posted on the list, ppc750 @800Mhz machine): > > - without the patches: > > real 4m 38.27s > user 0m 0.00s > sys 2m 18.90s > > - with the patches: > > real 0m 0.10s > user 0m 0.00s > sys 0m 0.05s > > Thanks, > tavi -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Octavian Purdila a écrit : > RFC patches are attached. > > Another possible approach: add an interface flag and use it to decide whether > we want per interface sysctl entries or not. > Hmm, could we speedup sysctl instead, adding rbtree or something ? -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sunday 25 October 2009 23:37:19 you wrote: > Octavian Purdila a écrit : > > RFC patches are attached. > > > > Another possible approach: add an interface flag and use it to decide > > whether we want per interface sysctl entries or not. > > Hmm, could we speedup sysctl instead, adding rbtree or something ? > Very good point, I think this is the best solution for people using a moderately high number of interfaces (a few thousand). But for really large setups there is another issue: memory consumption. In fact, in order to be able to scale to 128K interfaces and still have a significant amount of memory available to applications we also had to disable sysfs and #ifdef CONFIG_SYSFS struct device from net_device. I would also argue that when you have such a large number of interfaces you don't need to change setting on a per interface basis. Or at least this is our case :) and I suspect that the case with a large number of PPP interfaces is similar. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Monday 26 October 2009 00:21:48 Octavian Purdila wrote: > Very good point, I think this is the best solution for people using a > moderately high number of interfaces (a few thousand). > > But for really large setups there is another issue: memory consumption. In > fact, in order to be able to scale to 128K interfaces and still have a > significant amount of memory available to applications we also had to > disable sysfs and #ifdef CONFIG_SYSFS struct device from net_device. > > I would also argue that when you have such a large number of interfaces you > don't need to change setting on a per interface basis. Or at least this is > our case :) and I suspect that the case with a large number of PPP > interfaces is similar. I will add also, sysctl -a (over busybox) on pppoe with 2k interfaces takes ages to complete. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sun, 25 Oct 2009 19:54:49 +0200 Octavian Purdila <opurdila@ixiacom.com> wrote: > > RFC patches are attached. > > Another possible approach: add an interface flag and use it to decide whether > we want per interface sysctl entries or not. > > Benchmarks for creating 1000 interface (with the ndst module previously posted > on the list, ppc750 @800Mhz machine): > > - without the patches: > > real 4m 38.27s > user 0m 0.00s > sys 2m 18.90s > > - with the patches: > > real 0m 0.10s > user 0m 0.00s > sys 0m 0.05s > > Thanks, > tavi I would rather optimize the algorithm than give up and make it not available. It should be possible to do better by just using some better programming.
On Monday 26 October 2009 00:21:48 Octavian Purdila wrote: > On Sunday 25 October 2009 23:37:19 you wrote: > > Octavian Purdila a écrit : > > > RFC patches are attached. > > > > > > Another possible approach: add an interface flag and use it to decide > > > whether we want per interface sysctl entries or not. > > > > Hmm, could we speedup sysctl instead, adding rbtree or something ? > > Very good point, I think this is the best solution for people using a > moderately high number of interfaces (a few thousand). > > But for really large setups there is another issue: memory consumption. In > fact, in order to be able to scale to 128K interfaces and still have a > significant amount of memory available to applications we also had to > disable sysfs and #ifdef CONFIG_SYSFS struct device from net_device. > > I would also argue that when you have such a large number of interfaces you > don't need to change setting on a per interface basis. Or at least this is > our case :) and I suspect that the case with a large number of PPP > interfaces is similar. > Another possible approach: shared settings for an interface group. If you have a large number of interfaces of the same type it would be nice if you could change some setting for the whole group instead of globally or individually. Is this approach feasible anyway? Or I'm talking rubbish. Cosmin. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
I test it on pppoe with 1k customers. It works flawlessly. When there is problem on network and i have massive users disconnect and then login, the bottleneck is in lock somewhere in creation of sysctl(according perf). PPPoE after 200-300 interfaces will start dying, and connection rate will drop to 20-50 customers per minute, load average will jump to 70-100 (i guess pppd processes waiting their turn). With this patch i am able to sustain 200-300 customers / minute login rate and perftop is "clear" now. Definitely this option is optional, and doesn't cut any functionality by default, just giving more choice. And for PPP (pppoe/pptp) NAS it is very useful. On Sunday 25 October 2009 19:54:49 Octavian Purdila wrote: > RFC patches are attached. > > Another possible approach: add an interface flag and use it to decide > whether we want per interface sysctl entries or not. > > Benchmarks for creating 1000 interface (with the ndst module previously > posted on the list, ppc750 @800Mhz machine): > > - without the patches: > > real 4m 38.27s > user 0m 0.00s > sys 2m 18.90s > > - with the patches: > > real 0m 0.10s > user 0m 0.00s > sys 0m 0.05s > > Thanks, > tavi -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
net: CONFIG_NET_SYSCTL_DEV: make per interface dev_snmp6 proc entries optional Use same CONFIG_NET_SYSCTL_DEV config option (we should probably rename it to a better name) to enable/disable per interface dev_snmp6 proc entries. --- //packages/linux_2.6.31/rc7/src/include/net/ipv6.h +++ //packages/linux_2.6.31/rc7/src/include/net/ipv6.h @@ -604,8 +604,14 @@ extern void udplite6_proc_exit(void); extern int ipv6_misc_proc_init(void); extern void ipv6_misc_proc_exit(void); + +#ifdef CONFIG_NET_SYSCTL_DEV extern int snmp6_register_dev(struct inet6_dev *idev); extern int snmp6_unregister_dev(struct inet6_dev *idev); +#else +static inline int snmp6_register_dev(struct inet6_dev *idev) { return 0; } +static inline int snmp6_unregister_dev(struct inet6_dev *idev) { return 0; } +#endif #else static inline int ac6_proc_init(struct net *net) { return 0; } --- //packages/linux_2.6.31/rc7/src/net/ipv6/proc.c +++ //packages/linux_2.6.31/rc7/src/net/ipv6/proc.c @@ -232,6 +232,7 @@ .release = single_release, }; +#ifdef CONFIG_NET_SYSCTL_DEV int snmp6_register_dev(struct inet6_dev *idev) { struct proc_dir_entry *p; @@ -266,6 +267,7 @@ idev->stats.proc_dir_entry = NULL; return 0; } +#endif static int ipv6_proc_init_net(struct net *net) {