Submitter | Neil Horman |
---|---|

Date | Oct. 1, 2008, 6:08 p.m. |

Message ID | <20081001180846.GB3566@hmsreliant.think-freely.org> |

Download | mbox | patch |

Permalink | /patch/2291/ |

State | RFC |

Delegated to: | David Miller |

Headers | show |

## Comments

On Wed, 1 Oct 2008, Neil Horman wrote: > Hey all- > Since Eric mentioned the use of statistical analysis to determine if > hash chains were growing improperly, I thought I would take a stab at such an > approach. I'm no statistics expert, but it would seem to me that simply > computing the standard deviation of all the non-zero chain lengths would give a > good guide point to determine if we needed to invalidate our hash table. I've > written the below implementation. I've not tested it (I'll be doing that here > shortly for the next few days), but I wanted to post it to get feedback from you > all on the subject. The highlights are: > > 1) We add a counter to rt_hash_bucket structure to track the length of each > individual chain. I realize this is sub-optimal, as it adds potentially lots of > memory to hash table as a whole (4 bytes * number of hash buckets). I'm afraid > I've not come up with a better way to track that yet. We also track the total > number of route entries and the number of non-zero length chains. Lastly we > also maintain a global maximum chain length which defines the longest chain we > will tolerate in the route table. This patch defines it as the mean chain > length plus one standard deviation. I believe the general rule of thumb for something like this is at least two standard deviations. For a normal distribution, one standard deviation covers about 68 % of the sample universe, while two standard deviations covers about 95 % (three standard deviations covers 99.73 %). See the Wikipedia entry: http://en.wikipedia.org/wiki/Standard_deviation -Bill -- 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

Bill Fink a écrit : > On Wed, 1 Oct 2008, Neil Horman wrote: > >> Hey all- >> Since Eric mentioned the use of statistical analysis to determine if >> hash chains were growing improperly, I thought I would take a stab at such an >> approach. I'm no statistics expert, but it would seem to me that simply >> computing the standard deviation of all the non-zero chain lengths would give a >> good guide point to determine if we needed to invalidate our hash table. I've >> written the below implementation. I've not tested it (I'll be doing that here >> shortly for the next few days), but I wanted to post it to get feedback from you >> all on the subject. The highlights are: >> >> 1) We add a counter to rt_hash_bucket structure to track the length of each >> individual chain. I realize this is sub-optimal, as it adds potentially lots of >> memory to hash table as a whole (4 bytes * number of hash buckets). I'm afraid >> I've not come up with a better way to track that yet. We also track the total >> number of route entries and the number of non-zero length chains. Lastly we >> also maintain a global maximum chain length which defines the longest chain we >> will tolerate in the route table. This patch defines it as the mean chain >> length plus one standard deviation. > > I believe the general rule of thumb for something like this is at > least two standard deviations. For a normal distribution, one standard > deviation covers about 68 % of the sample universe, while two standard > deviations covers about 95 % (three standard deviations covers 99.73 %). > See the Wikipedia entry: > > http://en.wikipedia.org/wiki/Standard_deviation > Thanks Bill for the pointer, this is the trick. I believe we should target "4σ 99.993666% " case. But we dont need to really compute Standard deviation at runtime, only find an (upper) approximation of it. For elasticity=4 and 512*1024 samples (mean < 4), I guess 4σ can be approximated by 20 or something. Thank you -- 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

Eric Dumazet a écrit : > Bill Fink a écrit : >> On Wed, 1 Oct 2008, Neil Horman wrote: >> >>> Hey all- >>> Since Eric mentioned the use of statistical analysis to determine if >>> hash chains were growing improperly, I thought I would take a stab at >>> such an >>> approach. I'm no statistics expert, but it would seem to me that simply >>> computing the standard deviation of all the non-zero chain lengths >>> would give a >>> good guide point to determine if we needed to invalidate our hash >>> table. I've >>> written the below implementation. I've not tested it (I'll be doing >>> that here >>> shortly for the next few days), but I wanted to post it to get >>> feedback from you >>> all on the subject. The highlights are: >>> >>> 1) We add a counter to rt_hash_bucket structure to track the length >>> of each >>> individual chain. I realize this is sub-optimal, as it adds >>> potentially lots of >>> memory to hash table as a whole (4 bytes * number of hash buckets). >>> I'm afraid >>> I've not come up with a better way to track that yet. We also track >>> the total >>> number of route entries and the number of non-zero length chains. >>> Lastly we >>> also maintain a global maximum chain length which defines the longest >>> chain we >>> will tolerate in the route table. This patch defines it as the mean >>> chain >>> length plus one standard deviation. >> >> I believe the general rule of thumb for something like this is at >> least two standard deviations. For a normal distribution, one standard >> deviation covers about 68 % of the sample universe, while two standard >> deviations covers about 95 % (three standard deviations covers 99.73 %). >> See the Wikipedia entry: >> >> http://en.wikipedia.org/wiki/Standard_deviation >> > > Thanks Bill for the pointer, this is the trick. > > I believe we should target "4σ 99.993666% " case. > > But we dont need to really compute Standard deviation at runtime, only > find an (upper) approximation of it. > > For elasticity=4 and 512*1024 samples (mean < 4), I guess 4σ can be > approximated by 20 or something. > Good estimation of Standard Deviation could be computed for free in rt_check_expire(). (each runs scans 20% of hash table with default tunables timeout & ip_rt_gc_interval) We could update 4σ estimation in this function, every minute (ip_rt_gc_interval) At softirq time we then can detect a particular hash chain is longer than 4σ estimation and trigger an appropriate action. This action is to : flush table, and while we do that, expand hash table if its current size is under ip_rt_max_size/elasticity... -- 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

Eric Dumazet a écrit : > Eric Dumazet a écrit : >> Bill Fink a écrit : >>> I believe the general rule of thumb for something like this is at >>> least two standard deviations. For a normal distribution, one standard >>> deviation covers about 68 % of the sample universe, while two standard >>> deviations covers about 95 % (three standard deviations covers 99.73 %). >>> See the Wikipedia entry: >>> >>> http://en.wikipedia.org/wiki/Standard_deviation >>> >> >> Thanks Bill for the pointer, this is the trick. >> >> I believe we should target "4σ 99.993666% " case. >> >> But we dont need to really compute Standard deviation at runtime, only >> find an (upper) approximation of it. >> >> For elasticity=4 and 512*1024 samples (mean < 4), I guess 4σ can be >> approximated by 20 or something. >> > > Good estimation of Standard Deviation could be computed for free > in rt_check_expire(). (each runs scans 20% of hash table with default > tunables timeout & ip_rt_gc_interval) > > We could update 4σ estimation in this function, every minute > (ip_rt_gc_interval) > > At softirq time we then can detect a particular hash chain is > longer than 4σ estimation and trigger an appropriate action. > > This action is to : flush table, and while we do that, expand hash table > if its current size is under ip_rt_max_size/elasticity... > I ran again my litle dirty program that reads /proc/kcore to explore rt hash table on a busy server and found : hptr=0xffff81082ec00000 hsize=524288 total=2299242 dst entries 4306 chains of length 0 23807 chains of length 1 60119 chains of length 2 95112 chains of length 3 106364 chains of length 4 92307 chains of length 5 65743 chains of length 6 39710 chains of length 7 20997 chains of length 8 15775 chains of length 9 39 chains of length 10 7 chains of length 11 2 chains of length 12 avg = 4.38546 sd = 1.94614 X = avg + 4*sd = 12.17 I tried various elasticity settings and found litle variations of X This is because lot of entries are in use (refcnt>1) and can not be freed, regardless of elasticity. -- 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 Thu, Oct 02, 2008 at 10:15:38AM +0200, Eric Dumazet wrote: > Eric Dumazet a écrit : >> Bill Fink a écrit : >>> On Wed, 1 Oct 2008, Neil Horman wrote: >>> >>>> Hey all- >>>> Since Eric mentioned the use of statistical analysis to determine if >>>> hash chains were growing improperly, I thought I would take a stab >>>> at such an >>>> approach. I'm no statistics expert, but it would seem to me that simply >>>> computing the standard deviation of all the non-zero chain lengths >>>> would give a >>>> good guide point to determine if we needed to invalidate our hash >>>> table. I've >>>> written the below implementation. I've not tested it (I'll be >>>> doing that here >>>> shortly for the next few days), but I wanted to post it to get >>>> feedback from you >>>> all on the subject. The highlights are: >>>> >>>> 1) We add a counter to rt_hash_bucket structure to track the length >>>> of each >>>> individual chain. I realize this is sub-optimal, as it adds >>>> potentially lots of >>>> memory to hash table as a whole (4 bytes * number of hash buckets). >>>> I'm afraid >>>> I've not come up with a better way to track that yet. We also >>>> track the total >>>> number of route entries and the number of non-zero length chains. >>>> Lastly we >>>> also maintain a global maximum chain length which defines the >>>> longest chain we >>>> will tolerate in the route table. This patch defines it as the >>>> mean chain >>>> length plus one standard deviation. >>> >>> I believe the general rule of thumb for something like this is at >>> least two standard deviations. For a normal distribution, one standard >>> deviation covers about 68 % of the sample universe, while two standard >>> deviations covers about 95 % (three standard deviations covers 99.73 %). >>> See the Wikipedia entry: >>> >>> http://en.wikipedia.org/wiki/Standard_deviation >>> >> >> Thanks Bill for the pointer, this is the trick. >> >> I believe we should target "4σ 99.993666% " case. >> >> But we dont need to really compute Standard deviation at runtime, only >> find an (upper) approximation of it. >> >> For elasticity=4 and 512*1024 samples (mean < 4), I guess 4σ can be >> approximated by 20 or something. >> > > Good estimation of Standard Deviation could be computed for free > in rt_check_expire(). (each runs scans 20% of hash table with default > tunables timeout & ip_rt_gc_interval) > > We could update 4σ estimation in this function, every minute (ip_rt_gc_interval) > > At softirq time we then can detect a particular hash chain is > longer than 4σ estimation and trigger an appropriate action. > > This action is to : flush table, and while we do that, expand hash table > if its current size is under ip_rt_max_size/elasticity... > That is a good idea. I'm still testing the last patch I posted, and Its going pretty well (I did find a few bugs, so I'll be reposting when I'm done). Doing the update on demand when seem to be pretty quick, but I'll rewrite to gather stats on how long it take specifically, and then rewrite/compare to your suggested implementation above. It seems like the above would save space in the hash table, as I could eliminate the chain_length counter per hash bucket. On the down side though, I don't know if a 1 minute garbage collection interval would allow too much of a window to grow a chain in (i.e. we could check chain lengths against a too small standard deviation value, and trigger a cache rebuild unnecessecarily). Just thinking out loud, I've not got any evidence to support or contradict that Best Neil > > > > -- 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

## Patch

diff --git a/net/ipv4/route.c b/net/ipv4/route.c index 6ee5354..8e59900 100644 --- a/net/ipv4/route.c +++ b/net/ipv4/route.c @@ -145,6 +145,7 @@ static struct dst_entry *ipv4_negative_advice(struct dst_entry *dst); static void ipv4_link_failure(struct sk_buff *skb); static void ip_rt_update_pmtu(struct dst_entry *dst, u32 mtu); static int rt_garbage_collect(struct dst_ops *ops); +static void rt_emergency_hash_rebuild(struct net *net); static struct dst_ops ipv4_dst_ops = { @@ -200,7 +201,14 @@ const __u8 ip_tos2prio[16] = { struct rt_hash_bucket { struct rtable *chain; + atomic_t chain_length; }; + +atomic_t rt_hash_total_count; +atomic_t rt_hash_nz_count; + +static int rt_chain_length_max; + #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \ defined(CONFIG_PROVE_LOCKING) /* @@ -601,6 +609,53 @@ static inline int ip_rt_proc_init(void) } #endif /* CONFIG_PROC_FS */ +static void rt_hash_sd_update(void) +{ + int temp, i; + unsigned long long sd; + int average = atomic_read(&rt_hash_total_count); + int nzcount = atomic_read(&rt_hash_nz_count); + + average = DIV_ROUND_UP(average, nzcount); + + sd = 0; + for (i = 0; i < (1 << rt_hash_log); i++) { + temp = atomic_read(&rt_hash_table[i].chain_length); + if (unlikely(temp)) + sd += (temp-average)^2; + } + + sd = DIV_ROUND_UP(sd, nzcount); + + /* + * Note: 46340 squared is 0x7fffffff + */ + for (i = 1; i < 46340; i++) + if ((i^2) > sd) + break; + + /* + * The maximum allowable length of a hash + * chain is the average plus one standard + * deviation + */ + rt_chain_length_max = average + sd; +} + +static inline void rt_hash_sd_dec(int hash) +{ + atomic_dec(&rt_hash_table[hash].chain_length); + if (atomic_dec_and_test(&rt_hash_total_count)) + atomic_dec(&rt_hash_nz_count); +} + +static inline void rt_hash_sd_inc(int hash) +{ + atomic_inc(&rt_hash_table[hash].chain_length); + if (atomic_inc_return(&rt_hash_total_count) == 1) + atomic_inc(&rt_hash_nz_count); +} + static inline void rt_free(struct rtable *rt) { call_rcu_bh(&rt->u.dst.rcu_head, dst_rcu_free); @@ -731,6 +786,7 @@ static void rt_do_flush(int process_context) } else { *prev = next; rt_free(p); + rt_hash_sd_dec(i); } } } @@ -744,7 +800,9 @@ static void rt_do_flush(int process_context) for (; rth != tail; rth = next) { next = rth->u.dst.rt_next; rt_free(rth); + rt_hash_sd_dec(i); } + } } @@ -777,6 +835,7 @@ static void rt_check_expire(void) if (rt_is_expired(rth)) { *rthp = rth->u.dst.rt_next; rt_free(rth); + rt_hash_sd_dec(i); continue; } if (rth->u.dst.expires) { @@ -795,6 +854,7 @@ static void rt_check_expire(void) /* Cleanup aged off entries. */ *rthp = rth->u.dst.rt_next; rt_free(rth); + rt_hash_sd_dec(i); } spin_unlock_bh(rt_hash_lock_addr(i)); } @@ -927,6 +987,7 @@ static int rt_garbage_collect(struct dst_ops *ops) } *rthp = rth->u.dst.rt_next; rt_free(rth); + rt_hash_sd_dec(i); goal--; } spin_unlock_bh(rt_hash_lock_addr(k)); @@ -991,7 +1052,6 @@ static int rt_intern_hash(unsigned hash, struct rtable *rt, struct rtable **rp) int attempts = !in_softirq(); restart: - chain_length = 0; min_score = ~(u32)0; cand = NULL; candp = NULL; @@ -1004,6 +1064,7 @@ restart: if (rt_is_expired(rth)) { *rthp = rth->u.dst.rt_next; rt_free(rth); + rt_hash_sd_dec(hash); continue; } if (compare_keys(&rth->fl, &rt->fl) && compare_netns(rth, rt)) { @@ -1040,11 +1101,11 @@ restart: } } - chain_length++; rthp = &rth->u.dst.rt_next; } + chain_length = atomic_read(&rt_hash_table[hash].chain_length); if (cand) { /* ip_rt_gc_elasticity used to be average length of chain * length, when exceeded gc becomes really aggressive. @@ -1055,6 +1116,23 @@ restart: if (chain_length > ip_rt_gc_elasticity) { *candp = cand->u.dst.rt_next; rt_free(cand); + rt_hash_sd_dec(hash); + } + } else { + if (chain_length && (chain_length > rt_chain_length_max)) { + /* + * This chain is longer than our computed + * max allowable length, lets recompute to + * be sure its still up to date + */ + rt_hash_sd_update(); + if (chain_length > rt_chain_length_max) { + /* + * We're still to long, do an emergency rebuild + */ + rt_emergency_hash_rebuild(dev_net(rt->u.dst.dev)); + } + } } @@ -1106,6 +1184,7 @@ restart: #endif rt_hash_table[hash].chain = rt; spin_unlock_bh(rt_hash_lock_addr(hash)); + rt_hash_sd_inc(hash); *rp = rt; return 0; } @@ -1180,6 +1259,7 @@ static void rt_del(unsigned hash, struct rtable *rt) if (aux == rt || rt_is_expired(aux)) { *rthp = aux->u.dst.rt_next; rt_free(aux); + rt_hash_sd_dec(hash); continue; } rthp = &aux->u.dst.rt_next; @@ -2946,6 +3026,24 @@ static void rt_secret_reschedule(int old) rtnl_unlock(); } +static void rt_secret_rebuild_oneshot(struct net *net) { + del_timer_sync(&net->ipv4.rt_secret_timer); + rt_cache_invalidate(net); + if (ip_rt_secret_interval) { + net->ipv4.rt_secret_timer.expires += ip_rt_secret_interval; + add_timer(&net->ipv4.rt_secret_timer); + } +} + +static void rt_emergency_hash_rebuild(struct net *net) { + if (net_ratelimit()) { + printk(KERN_WARNING "Route hash chain too long!\n"); + printk(KERN_WARNING "Adjust your secret_interval!\n"); + } + + rt_secret_rebuild_oneshot(net); +} + static int ipv4_sysctl_rt_secret_interval(ctl_table *ctl, int write, struct file *filp, void __user *buffer, size_t *lenp,

`Hey all- Since Eric mentioned the use of statistical analysis to determine if hash chains were growing improperly, I thought I would take a stab at such an approach. I'm no statistics expert, but it would seem to me that simply computing the standard deviation of all the non-zero chain lengths would give a good guide point to determine if we needed to invalidate our hash table. I've written the below implementation. I've not tested it (I'll be doing that here shortly for the next few days), but I wanted to post it to get feedback from you all on the subject. The highlights are: 1) We add a counter to rt_hash_bucket structure to track the length of each individual chain. I realize this is sub-optimal, as it adds potentially lots of memory to hash table as a whole (4 bytes * number of hash buckets). I'm afraid I've not come up with a better way to track that yet. We also track the total number of route entries and the number of non-zero length chains. Lastly we also maintain a global maximum chain length which defines the longest chain we will tolerate in the route table. This patch defines it as the mean chain length plus one standard deviation. 2) The above new counters are updated as we add/delete entries from the hash table 3) in rt_intern_hash if we don't find an entry to delete while walking a given chain that is over its elasticity, we check the chain length against the maximum length, which we defined in 1. If it exceeds the max length, we recompute the mean and standard deviation value, and check the length again (see rt_hash_sd_update). If the length is still longer than the new max length, we execute an emergency hash rebuild. Again, I've not started testing it, but if this works, we should be able to eliminate the need for periodic cache rebuilds entirely, replacing it with on demand detection. I could see a usefulness for add a control to tweak the multiple count of standard deviations allowed from the mean, for environments where route entries just happened to hash to the same bucket, but I figure, let me get this working properly first. I'll be doing basic testing over the next few days. Any thoughts & comments are appreciated, as well as testing in more strenuous, high route entry/high throughput environments. Thanks & Regards Neil Signed-off-by: Neil Horman <nhorman@tuxdriver.com> route.c | 102 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 100 insertions(+), 2 deletions(-)`