| Message ID | 1254000675-8327-1-git-send-email-iler.ml@gmail.com |
|---|---|
| State | Rejected, archived |
| Delegated to: | David Miller |
| Headers | show |
Yakov Lerner a écrit : > /proc/net/tcp does 20,000 sockets in 60-80 milliseconds, with this patch. > > The overhead was in tcp_seq_start(). See analysis (3) below. > The patch is against Linus git tree (1). The patch is small. > > ------------ ----------- ------------------------------------ > Before patch After patch 20,000 sockets (10,000 tw + 10,000 estab)(2) > ------------ ----------- ------------------------------------ > 6 sec 0.06 sec dd bs=1k if=/proc/net/tcp >/dev/null > 1.5 sec 0.06 sec dd bs=4k if=/proc/net/tcp >/dev/null > > 1.9 sec 0.16 sec netstat -4ant >/dev/null > ------------ ----------- ------------------------------------ > > This is ~ x25 improvement. > The new time is not dependent on read blockize. > Speed of netstat, naturally, improves, too; both -4 and -6. > /proc/net/tcp6 does 20,000 sockets in 100 millisec. > > (1) against git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git > > (2) Used 'manysock' utility to stress system with large number of sockets: > "manysock 10000 10000" - 10,000 tw + 10,000 estab ip4 sockets. > "manysock -6 10000 10000" - 10,000 tw + 10,000 estab ip6 sockets. > Found at http://ilerner.3b1.org/manysock/manysock.c > > (3) Algorithmic analysis. > Old algorithm. > > During 'cat </proc/net/tcp', tcp_seq_start() is called O(numsockets) times (4). > On average, every call to tcp_seq_start() scans half the whole hashtable. Ouch. > This is O(numsockets * hashsize). 95-99% of 'cat </proc/net/tcp' is spent in > tcp_seq_start()->tcp_get_idx. This overhead is eliminated by new algorithm, > which is O(numsockets + hashsize). > > New algorithm. > > New algorithms is O(numsockets + hashsize). We jump to the right > hash bucket in tcp_seq_start(), without scanning half the hash. > To jump right to the hash bucket corresponding to *pos in tcp_seq_start(), > we reuse three pieces of state (st->num, st->bucket, st->sbucket) > as follows: > - we check that requested pos >= last seen pos (st->num), the typical case. > - if so, we jump to bucket st->bucket > - to arrive to the right item after beginning of st->bucket, we > keep in st->sbucket the position corresponding to the beginning of > bucket. > > (4) Explanation of O( numsockets * hashsize) of old algorithm. > > tcp_seq_start() is called once for every ~7 lines of netstat output > if readsize is 1kb, or once for every ~28 lines if readsize >= 4kb. > Since record length of /proc/net/tcp records is 150 bytes, formula for > number of calls to tcp_seq_start() is > (numsockets * 150 / min(4096,readsize)). > Netstat uses 4kb readsize (newer versions), or 1kb (older versions). > Note that speed of old algorithm does not improve above 4kb blocksize. > > Speed of the new algorithm does not depend on blocksize. > > Speed of the new algorithm does not perceptibly depend on hashsize (which > depends on ramsize). Speed of old algorithm drops with bigger hashsize. > > (5) Reporting order. > > Reporting order is exactly same as before if hash does not change underfoot. > When hash elements come and go during report, reporting order will be > same as that of tcpdiag. > > Signed-off-by: Yakov Lerner <iler.ml@gmail.com> > --- > net/ipv4/tcp_ipv4.c | 26 ++++++++++++++++++++++++-- > 1 files changed, 24 insertions(+), 2 deletions(-) > > diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c > index 7cda24b..7d9421a 100644 > --- a/net/ipv4/tcp_ipv4.c > +++ b/net/ipv4/tcp_ipv4.c > @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st) > hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain); > } > > -static void *established_get_first(struct seq_file *seq) > +static void *established_get_first_after(struct seq_file *seq, int bucket) > { > struct tcp_iter_state *st = seq->private; > struct net *net = seq_file_net(seq); > void *rc = NULL; > > - for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) { > + for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size; > + ++st->bucket) { > struct sock *sk; > struct hlist_nulls_node *node; > struct inet_timewait_sock *tw; > @@ -2036,6 +2037,11 @@ out: > return rc; > } > > +static void *established_get_first(struct seq_file *seq) > +{ > + return established_get_first_after(seq, 0); > +} > + > static void *established_get_next(struct seq_file *seq, void *cur) > { > struct sock *sk = cur; > @@ -2045,6 +2051,7 @@ static void *established_get_next(struct seq_file *seq, void *cur) > struct net *net = seq_file_net(seq); > > ++st->num; > + st->sbucket = st->num; Hello Yakov Intention of your patch is very good, but not currently working. It seems you believe there is at most one entry per hash slot or something like that Please reboot your test machine with "thash_entries=4096" so that tcp hash size is 4096, and try to fill 20000 tcp sockets with a test program. then : # ss | wc -l 20001 (ok) # cat /proc/net/tcp | wc -l 22160 (not quite correct ...) # netstat -tn | wc -l <never ends> # dd if=/proc/net/tcp ibs=1024 | wc -l <never ends> Please send your next patch on netdev@vger.kernel.org , DaveM only , were netdev people are reviewing netdev patches, there is no need include other people for first submissions. Thank you #include <sys/types.h> #include <sys/socket.h> #include <netinet/in.h> #include <string.h> int fdlisten; main() { int i; struct sockaddr_in sockaddr; fdlisten = socket(AF_INET, SOCK_STREAM, 0); memset(&sockaddr, 0, sizeof(sockaddr)); sockaddr.sin_family = AF_INET; sockaddr.sin_port = htons(2222); if (bind(fdlisten, (struct sockaddr *)&sockaddr, sizeof(sockaddr))== -1) { perror("bind"); return 1; } if (listen(fdlisten, 10)== -1) { perror("listen"); return 1; } if (fork() == 0) { while (1) { socklen_t len = sizeof(sockaddr); int newfd = accept(fdlisten, (struct sockaddr *)&sockaddr, &len); } } for (i = 0 ; i < 10000; i++) { int fd = socket(AF_INET, SOCK_STREAM, 0); if (fd == -1) { perror("socket"); break; } connect(fd, (struct sockaddr *)&sockaddr, sizeof(sockaddr)); } pause(); } -- 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, Sep 27, 2009 at 12:53, Eric Dumazet <eric.dumazet@gmail.com> wrote: > Yakov Lerner a écrit : >> /proc/net/tcp does 20,000 sockets in 60-80 milliseconds, with this patch. >> >> The overhead was in tcp_seq_start(). See analysis (3) below. >> The patch is against Linus git tree (1). The patch is small. >> >> ------------ ----------- ------------------------------------ >> Before patch After patch 20,000 sockets (10,000 tw + 10,000 estab)(2) >> ------------ ----------- ------------------------------------ >> 6 sec 0.06 sec dd bs=1k if=/proc/net/tcp >/dev/null >> 1.5 sec 0.06 sec dd bs=4k if=/proc/net/tcp >/dev/null >> >> 1.9 sec 0.16 sec netstat -4ant >/dev/null >> ------------ ----------- ------------------------------------ >> >> This is ~ x25 improvement. >> The new time is not dependent on read blockize. >> Speed of netstat, naturally, improves, too; both -4 and -6. >> /proc/net/tcp6 does 20,000 sockets in 100 millisec. >> >> (1) against git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git >> >> (2) Used 'manysock' utility to stress system with large number of sockets: >> "manysock 10000 10000" - 10,000 tw + 10,000 estab ip4 sockets. >> "manysock -6 10000 10000" - 10,000 tw + 10,000 estab ip6 sockets. >> Found at http://ilerner.3b1.org/manysock/manysock.c >> >> (3) Algorithmic analysis. >> Old algorithm. >> >> During 'cat </proc/net/tcp', tcp_seq_start() is called O(numsockets) times (4). >> On average, every call to tcp_seq_start() scans half the whole hashtable. Ouch. >> This is O(numsockets * hashsize). 95-99% of 'cat </proc/net/tcp' is spent in >> tcp_seq_start()->tcp_get_idx. This overhead is eliminated by new algorithm, >> which is O(numsockets + hashsize). >> >> New algorithm. >> >> New algorithms is O(numsockets + hashsize). We jump to the right >> hash bucket in tcp_seq_start(), without scanning half the hash. >> To jump right to the hash bucket corresponding to *pos in tcp_seq_start(), >> we reuse three pieces of state (st->num, st->bucket, st->sbucket) >> as follows: >> - we check that requested pos >= last seen pos (st->num), the typical case. >> - if so, we jump to bucket st->bucket >> - to arrive to the right item after beginning of st->bucket, we >> keep in st->sbucket the position corresponding to the beginning of >> bucket. >> >> (4) Explanation of O( numsockets * hashsize) of old algorithm. >> >> tcp_seq_start() is called once for every ~7 lines of netstat output >> if readsize is 1kb, or once for every ~28 lines if readsize >= 4kb. >> Since record length of /proc/net/tcp records is 150 bytes, formula for >> number of calls to tcp_seq_start() is >> (numsockets * 150 / min(4096,readsize)). >> Netstat uses 4kb readsize (newer versions), or 1kb (older versions). >> Note that speed of old algorithm does not improve above 4kb blocksize. >> >> Speed of the new algorithm does not depend on blocksize. >> >> Speed of the new algorithm does not perceptibly depend on hashsize (which >> depends on ramsize). Speed of old algorithm drops with bigger hashsize. >> >> (5) Reporting order. >> >> Reporting order is exactly same as before if hash does not change underfoot. >> When hash elements come and go during report, reporting order will be >> same as that of tcpdiag. >> >> Signed-off-by: Yakov Lerner <iler.ml@gmail.com> >> --- >> net/ipv4/tcp_ipv4.c | 26 ++++++++++++++++++++++++-- >> 1 files changed, 24 insertions(+), 2 deletions(-) >> >> diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c >> index 7cda24b..7d9421a 100644 >> --- a/net/ipv4/tcp_ipv4.c >> +++ b/net/ipv4/tcp_ipv4.c >> @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st) >> hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain); >> } >> >> -static void *established_get_first(struct seq_file *seq) >> +static void *established_get_first_after(struct seq_file *seq, int bucket) >> { >> struct tcp_iter_state *st = seq->private; >> struct net *net = seq_file_net(seq); >> void *rc = NULL; >> >> - for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) { >> + for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size; >> + ++st->bucket) { >> struct sock *sk; >> struct hlist_nulls_node *node; >> struct inet_timewait_sock *tw; >> @@ -2036,6 +2037,11 @@ out: >> return rc; >> } >> >> +static void *established_get_first(struct seq_file *seq) >> +{ >> + return established_get_first_after(seq, 0); >> +} >> + >> static void *established_get_next(struct seq_file *seq, void *cur) >> { >> struct sock *sk = cur; >> @@ -2045,6 +2051,7 @@ static void *established_get_next(struct seq_file *seq, void *cur) >> struct net *net = seq_file_net(seq); >> >> ++st->num; >> + st->sbucket = st->num; > > Hello Yakov > > Intention of your patch is very good, but not currently working. > > It seems you believe there is at most one entry per hash slot or something like that > > Please reboot your test machine with "thash_entries=4096" so that tcp hash > size is 4096, and try to fill 20000 tcp sockets with a test program. > > then : > > # ss | wc -l > 20001 > (ok) > > # cat /proc/net/tcp | wc -l > 22160 > (not quite correct ...) > > # netstat -tn | wc -l > <never ends> > > > # dd if=/proc/net/tcp ibs=1024 | wc -l > <never ends> > > > Please send your next patch on netdev@vger.kernel.org , DaveM only , were netdev people > are reviewing netdev patches, there is no need include other people for first submissions. > > Thank you > > > #include <sys/types.h> > #include <sys/socket.h> > #include <netinet/in.h> > #include <string.h> > int fdlisten; > main() > { > int i; > struct sockaddr_in sockaddr; > > fdlisten = socket(AF_INET, SOCK_STREAM, 0); > memset(&sockaddr, 0, sizeof(sockaddr)); > sockaddr.sin_family = AF_INET; > sockaddr.sin_port = htons(2222); > if (bind(fdlisten, (struct sockaddr *)&sockaddr, sizeof(sockaddr))== -1) { > perror("bind"); > return 1; > } > if (listen(fdlisten, 10)== -1) { > perror("listen"); > return 1; > } > if (fork() == 0) { > while (1) { > socklen_t len = sizeof(sockaddr); > int newfd = accept(fdlisten, (struct sockaddr *)&sockaddr, &len); > } > } > for (i = 0 ; i < 10000; i++) { > int fd = socket(AF_INET, SOCK_STREAM, 0); > if (fd == -1) { > perror("socket"); > break; > } > connect(fd, (struct sockaddr *)&sockaddr, sizeof(sockaddr)); > } > pause(); > } > Hello Eric, I found the problem, thanks. I'll re-send after testing. In the meantime, I'd like to ask you whether it makes sense to add the /proc/net entry, to switch between "old way" and "new way". The switch would allow quick compare/test between new way and old way not only by line count, but by full contents, without reboot. Yakov -- 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
Yakov Lerner a écrit : > On Sun, Sep 27, 2009 at 12:53, Eric Dumazet <eric.dumazet@gmail.com> wrote: >> Yakov Lerner a écrit : >>> /proc/net/tcp does 20,000 sockets in 60-80 milliseconds, with this patch. >>> >>> The overhead was in tcp_seq_start(). See analysis (3) below. >>> The patch is against Linus git tree (1). The patch is small. >>> >>> ------------ ----------- ------------------------------------ >>> Before patch After patch 20,000 sockets (10,000 tw + 10,000 estab)(2) >>> ------------ ----------- ------------------------------------ >>> 6 sec 0.06 sec dd bs=1k if=/proc/net/tcp >/dev/null >>> 1.5 sec 0.06 sec dd bs=4k if=/proc/net/tcp >/dev/null >>> >>> 1.9 sec 0.16 sec netstat -4ant >/dev/null >>> ------------ ----------- ------------------------------------ >>> >>> This is ~ x25 improvement. >>> The new time is not dependent on read blockize. >>> Speed of netstat, naturally, improves, too; both -4 and -6. >>> /proc/net/tcp6 does 20,000 sockets in 100 millisec. >>> >>> (1) against git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git >>> >>> (2) Used 'manysock' utility to stress system with large number of sockets: >>> "manysock 10000 10000" - 10,000 tw + 10,000 estab ip4 sockets. >>> "manysock -6 10000 10000" - 10,000 tw + 10,000 estab ip6 sockets. >>> Found at http://ilerner.3b1.org/manysock/manysock.c >>> >>> (3) Algorithmic analysis. >>> Old algorithm. >>> >>> During 'cat </proc/net/tcp', tcp_seq_start() is called O(numsockets) times (4). >>> On average, every call to tcp_seq_start() scans half the whole hashtable. Ouch. >>> This is O(numsockets * hashsize). 95-99% of 'cat </proc/net/tcp' is spent in >>> tcp_seq_start()->tcp_get_idx. This overhead is eliminated by new algorithm, >>> which is O(numsockets + hashsize). >>> >>> New algorithm. >>> >>> New algorithms is O(numsockets + hashsize). We jump to the right >>> hash bucket in tcp_seq_start(), without scanning half the hash. >>> To jump right to the hash bucket corresponding to *pos in tcp_seq_start(), >>> we reuse three pieces of state (st->num, st->bucket, st->sbucket) >>> as follows: >>> - we check that requested pos >= last seen pos (st->num), the typical case. >>> - if so, we jump to bucket st->bucket >>> - to arrive to the right item after beginning of st->bucket, we >>> keep in st->sbucket the position corresponding to the beginning of >>> bucket. >>> >>> (4) Explanation of O( numsockets * hashsize) of old algorithm. >>> >>> tcp_seq_start() is called once for every ~7 lines of netstat output >>> if readsize is 1kb, or once for every ~28 lines if readsize >= 4kb. >>> Since record length of /proc/net/tcp records is 150 bytes, formula for >>> number of calls to tcp_seq_start() is >>> (numsockets * 150 / min(4096,readsize)). >>> Netstat uses 4kb readsize (newer versions), or 1kb (older versions). >>> Note that speed of old algorithm does not improve above 4kb blocksize. >>> >>> Speed of the new algorithm does not depend on blocksize. >>> >>> Speed of the new algorithm does not perceptibly depend on hashsize (which >>> depends on ramsize). Speed of old algorithm drops with bigger hashsize. >>> >>> (5) Reporting order. >>> >>> Reporting order is exactly same as before if hash does not change underfoot. >>> When hash elements come and go during report, reporting order will be >>> same as that of tcpdiag. >>> >>> Signed-off-by: Yakov Lerner <iler.ml@gmail.com> >>> --- >>> net/ipv4/tcp_ipv4.c | 26 ++++++++++++++++++++++++-- >>> 1 files changed, 24 insertions(+), 2 deletions(-) >>> >>> diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c >>> index 7cda24b..7d9421a 100644 >>> --- a/net/ipv4/tcp_ipv4.c >>> +++ b/net/ipv4/tcp_ipv4.c >>> @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st) >>> hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain); >>> } >>> >>> -static void *established_get_first(struct seq_file *seq) >>> +static void *established_get_first_after(struct seq_file *seq, int bucket) >>> { >>> struct tcp_iter_state *st = seq->private; >>> struct net *net = seq_file_net(seq); >>> void *rc = NULL; >>> >>> - for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) { >>> + for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size; >>> + ++st->bucket) { >>> struct sock *sk; >>> struct hlist_nulls_node *node; >>> struct inet_timewait_sock *tw; >>> @@ -2036,6 +2037,11 @@ out: >>> return rc; >>> } >>> >>> +static void *established_get_first(struct seq_file *seq) >>> +{ >>> + return established_get_first_after(seq, 0); >>> +} >>> + >>> static void *established_get_next(struct seq_file *seq, void *cur) >>> { >>> struct sock *sk = cur; >>> @@ -2045,6 +2051,7 @@ static void *established_get_next(struct seq_file *seq, void *cur) >>> struct net *net = seq_file_net(seq); >>> >>> ++st->num; >>> + st->sbucket = st->num; >> Hello Yakov >> >> Intention of your patch is very good, but not currently working. >> >> It seems you believe there is at most one entry per hash slot or something like that >> >> Please reboot your test machine with "thash_entries=4096" so that tcp hash >> size is 4096, and try to fill 20000 tcp sockets with a test program. >> >> then : >> >> # ss | wc -l >> 20001 >> (ok) >> >> # cat /proc/net/tcp | wc -l >> 22160 >> (not quite correct ...) >> >> # netstat -tn | wc -l >> <never ends> >> >> >> # dd if=/proc/net/tcp ibs=1024 | wc -l >> <never ends> >> >> >> Please send your next patch on netdev@vger.kernel.org , DaveM only , were netdev people >> are reviewing netdev patches, there is no need include other people for first submissions. >> >> Thank you >> >> >> #include <sys/types.h> >> #include <sys/socket.h> >> #include <netinet/in.h> >> #include <string.h> >> int fdlisten; >> main() >> { >> int i; >> struct sockaddr_in sockaddr; >> >> fdlisten = socket(AF_INET, SOCK_STREAM, 0); >> memset(&sockaddr, 0, sizeof(sockaddr)); >> sockaddr.sin_family = AF_INET; >> sockaddr.sin_port = htons(2222); >> if (bind(fdlisten, (struct sockaddr *)&sockaddr, sizeof(sockaddr))== -1) { >> perror("bind"); >> return 1; >> } >> if (listen(fdlisten, 10)== -1) { >> perror("listen"); >> return 1; >> } >> if (fork() == 0) { >> while (1) { >> socklen_t len = sizeof(sockaddr); >> int newfd = accept(fdlisten, (struct sockaddr *)&sockaddr, &len); >> } >> } >> for (i = 0 ; i < 10000; i++) { >> int fd = socket(AF_INET, SOCK_STREAM, 0); >> if (fd == -1) { >> perror("socket"); >> break; >> } >> connect(fd, (struct sockaddr *)&sockaddr, sizeof(sockaddr)); >> } >> pause(); >> } >> > > Hello Eric, > > I found the problem, thanks. I'll re-send after testing. OK good ! > > In the meantime, I'd like to ask you whether it makes sense to > add the /proc/net entry, to switch between "old way" and "new way". > The switch would allow quick compare/test between new way and > old way not only by line count, but by full contents, without reboot. > Well, this switch wont be needed for patch validation, but it might help you to test your patch of course. Actually I found the error reading your patch, and I made a quick test to confirm my understanding :) See you tomorrow, its rather late here :) -- 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 Tue, 29 Sep 2009 00:20:07 +0200 Eric Dumazet <eric.dumazet@gmail.com> wrote: > Yakov Lerner a écrit : > > On Sun, Sep 27, 2009 at 12:53, Eric Dumazet <eric.dumazet@gmail.com> wrote: > >> Yakov Lerner a écrit : > >>> /proc/net/tcp does 20,000 sockets in 60-80 milliseconds, with this patch. > >>> > >>> The overhead was in tcp_seq_start(). See analysis (3) below. > >>> The patch is against Linus git tree (1). The patch is small. > >>> > >>> ------------ ----------- ------------------------------------ > >>> Before patch After patch 20,000 sockets (10,000 tw + 10,000 estab)(2) > >>> ------------ ----------- ------------------------------------ > >>> 6 sec 0.06 sec dd bs=1k if=/proc/net/tcp >/dev/null > >>> 1.5 sec 0.06 sec dd bs=4k if=/proc/net/tcp >/dev/null > >>> > >>> 1.9 sec 0.16 sec netstat -4ant >/dev/null > >>> ------------ ----------- ------------------------------------ > >>> > >>> This is ~ x25 improvement. > >>> The new time is not dependent on read blockize. > >>> Speed of netstat, naturally, improves, too; both -4 and -6. > >>> /proc/net/tcp6 does 20,000 sockets in 100 millisec. > >>> > >>> (1) against git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git > >>> > >>> (2) Used 'manysock' utility to stress system with large number of sockets: > >>> "manysock 10000 10000" - 10,000 tw + 10,000 estab ip4 sockets. > >>> "manysock -6 10000 10000" - 10,000 tw + 10,000 estab ip6 sockets. > >>> Found at http://ilerner.3b1.org/manysock/manysock.c > >>> > >>> (3) Algorithmic analysis. > >>> Old algorithm. > >>> > >>> During 'cat </proc/net/tcp', tcp_seq_start() is called O(numsockets) times (4). > >>> On average, every call to tcp_seq_start() scans half the whole hashtable. Ouch. > >>> This is O(numsockets * hashsize). 95-99% of 'cat </proc/net/tcp' is spent in > >>> tcp_seq_start()->tcp_get_idx. This overhead is eliminated by new algorithm, > >>> which is O(numsockets + hashsize). > >>> > >>> New algorithm. > >>> > >>> New algorithms is O(numsockets + hashsize). We jump to the right > >>> hash bucket in tcp_seq_start(), without scanning half the hash. > >>> To jump right to the hash bucket corresponding to *pos in tcp_seq_start(), > >>> we reuse three pieces of state (st->num, st->bucket, st->sbucket) > >>> as follows: > >>> - we check that requested pos >= last seen pos (st->num), the typical case. > >>> - if so, we jump to bucket st->bucket > >>> - to arrive to the right item after beginning of st->bucket, we > >>> keep in st->sbucket the position corresponding to the beginning of > >>> bucket. > >>> > >>> (4) Explanation of O( numsockets * hashsize) of old algorithm. > >>> > >>> tcp_seq_start() is called once for every ~7 lines of netstat output > >>> if readsize is 1kb, or once for every ~28 lines if readsize >= 4kb. > >>> Since record length of /proc/net/tcp records is 150 bytes, formula for > >>> number of calls to tcp_seq_start() is > >>> (numsockets * 150 / min(4096,readsize)). > >>> Netstat uses 4kb readsize (newer versions), or 1kb (older versions). > >>> Note that speed of old algorithm does not improve above 4kb blocksize. > >>> > >>> Speed of the new algorithm does not depend on blocksize. > >>> > >>> Speed of the new algorithm does not perceptibly depend on hashsize (which > >>> depends on ramsize). Speed of old algorithm drops with bigger hashsize. > >>> > >>> (5) Reporting order. > >>> > >>> Reporting order is exactly same as before if hash does not change underfoot. > >>> When hash elements come and go during report, reporting order will be > >>> same as that of tcpdiag. > >>> > >>> Signed-off-by: Yakov Lerner <iler.ml@gmail.com> Does the netlink interface used by ss command have the problem?
On Tue, Sep 29, 2009 at 02:24, Stephen Hemminger <shemminger@vyatta.com> wrote: > On Tue, 29 Sep 2009 00:20:07 +0200 > Eric Dumazet <eric.dumazet@gmail.com> wrote: > >> Yakov Lerner a écrit : >> > On Sun, Sep 27, 2009 at 12:53, Eric Dumazet <eric.dumazet@gmail.com> wrote: >> >> Yakov Lerner a écrit : >> >>> /proc/net/tcp does 20,000 sockets in 60-80 milliseconds, with this patch. >> >>> >> >>> The overhead was in tcp_seq_start(). See analysis (3) below. >> >>> The patch is against Linus git tree (1). The patch is small. >> >>> >> >>> ------------ ----------- ------------------------------------ >> >>> Before patch After patch 20,000 sockets (10,000 tw + 10,000 estab)(2) >> >>> ------------ ----------- ------------------------------------ >> >>> 6 sec 0.06 sec dd bs=1k if=/proc/net/tcp >/dev/null >> >>> 1.5 sec 0.06 sec dd bs=4k if=/proc/net/tcp >/dev/null >> >>> >> >>> 1.9 sec 0.16 sec netstat -4ant >/dev/null >> >>> ------------ ----------- ------------------------------------ >> >>> >> >>> This is ~ x25 improvement. >> >>> The new time is not dependent on read blockize. >> >>> Speed of netstat, naturally, improves, too; both -4 and -6. >> >>> /proc/net/tcp6 does 20,000 sockets in 100 millisec. >> >>> >> >>> (1) against git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git >> >>> >> >>> (2) Used 'manysock' utility to stress system with large number of sockets: >> >>> "manysock 10000 10000" - 10,000 tw + 10,000 estab ip4 sockets. >> >>> "manysock -6 10000 10000" - 10,000 tw + 10,000 estab ip6 sockets. >> >>> Found at http://ilerner.3b1.org/manysock/manysock.c >> >>> >> >>> (3) Algorithmic analysis. >> >>> Old algorithm. >> >>> >> >>> During 'cat </proc/net/tcp', tcp_seq_start() is called O(numsockets) times (4). >> >>> On average, every call to tcp_seq_start() scans half the whole hashtable. Ouch. >> >>> This is O(numsockets * hashsize). 95-99% of 'cat </proc/net/tcp' is spent in >> >>> tcp_seq_start()->tcp_get_idx. This overhead is eliminated by new algorithm, >> >>> which is O(numsockets + hashsize). >> >>> >> >>> New algorithm. >> >>> >> >>> New algorithms is O(numsockets + hashsize). We jump to the right >> >>> hash bucket in tcp_seq_start(), without scanning half the hash. >> >>> To jump right to the hash bucket corresponding to *pos in tcp_seq_start(), >> >>> we reuse three pieces of state (st->num, st->bucket, st->sbucket) >> >>> as follows: >> >>> - we check that requested pos >= last seen pos (st->num), the typical case. >> >>> - if so, we jump to bucket st->bucket >> >>> - to arrive to the right item after beginning of st->bucket, we >> >>> keep in st->sbucket the position corresponding to the beginning of >> >>> bucket. >> >>> >> >>> (4) Explanation of O( numsockets * hashsize) of old algorithm. >> >>> >> >>> tcp_seq_start() is called once for every ~7 lines of netstat output >> >>> if readsize is 1kb, or once for every ~28 lines if readsize >= 4kb. >> >>> Since record length of /proc/net/tcp records is 150 bytes, formula for >> >>> number of calls to tcp_seq_start() is >> >>> (numsockets * 150 / min(4096,readsize)). >> >>> Netstat uses 4kb readsize (newer versions), or 1kb (older versions). >> >>> Note that speed of old algorithm does not improve above 4kb blocksize. >> >>> >> >>> Speed of the new algorithm does not depend on blocksize. >> >>> >> >>> Speed of the new algorithm does not perceptibly depend on hashsize (which >> >>> depends on ramsize). Speed of old algorithm drops with bigger hashsize. >> >>> >> >>> (5) Reporting order. >> >>> >> >>> Reporting order is exactly same as before if hash does not change underfoot. >> >>> When hash elements come and go during report, reporting order will be >> >>> same as that of tcpdiag. >> >>> >> >>> Signed-off-by: Yakov Lerner <iler.ml@gmail.com> > > Does the netlink interface used by ss command have the problem? No. It's /proc/net/tcp that has fixable problem. Yakov -- 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
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c index 7cda24b..7d9421a 100644 --- a/net/ipv4/tcp_ipv4.c +++ b/net/ipv4/tcp_ipv4.c @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st) hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain); } -static void *established_get_first(struct seq_file *seq) +static void *established_get_first_after(struct seq_file *seq, int bucket) { struct tcp_iter_state *st = seq->private; struct net *net = seq_file_net(seq); void *rc = NULL; - for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) { + for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size; + ++st->bucket) { struct sock *sk; struct hlist_nulls_node *node; struct inet_timewait_sock *tw; @@ -2036,6 +2037,11 @@ out: return rc; } +static void *established_get_first(struct seq_file *seq) +{ + return established_get_first_after(seq, 0); +} + static void *established_get_next(struct seq_file *seq, void *cur) { struct sock *sk = cur; @@ -2045,6 +2051,7 @@ static void *established_get_next(struct seq_file *seq, void *cur) struct net *net = seq_file_net(seq); ++st->num; + st->sbucket = st->num; if (st->state == TCP_SEQ_STATE_TIME_WAIT) { tw = cur; @@ -2116,6 +2123,21 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos) static void *tcp_seq_start(struct seq_file *seq, loff_t *pos) { struct tcp_iter_state *st = seq->private; + + if (*pos && *pos >= st->sbucket && + (st->state == TCP_SEQ_STATE_ESTABLISHED || + st->state == TCP_SEQ_STATE_TIME_WAIT)) { + int nskip; + void *cur; + + st->num = st->sbucket; + st->state = TCP_SEQ_STATE_ESTABLISHED; + cur = established_get_first_after(seq, st->bucket); + for (nskip = *pos - st->sbucket; nskip > 0 && cur; --nskip) + cur = established_get_next(seq, cur); + return cur; + } + st->state = TCP_SEQ_STATE_LISTENING; st->num = 0; return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
/proc/net/tcp does 20,000 sockets in 60-80 milliseconds, with this patch. The overhead was in tcp_seq_start(). See analysis (3) below. The patch is against Linus git tree (1). The patch is small. ------------ ----------- ------------------------------------ Before patch After patch 20,000 sockets (10,000 tw + 10,000 estab)(2) ------------ ----------- ------------------------------------ 6 sec 0.06 sec dd bs=1k if=/proc/net/tcp >/dev/null 1.5 sec 0.06 sec dd bs=4k if=/proc/net/tcp >/dev/null 1.9 sec 0.16 sec netstat -4ant >/dev/null ------------ ----------- ------------------------------------ This is ~ x25 improvement. The new time is not dependent on read blockize. Speed of netstat, naturally, improves, too; both -4 and -6. /proc/net/tcp6 does 20,000 sockets in 100 millisec. (1) against git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git (2) Used 'manysock' utility to stress system with large number of sockets: "manysock 10000 10000" - 10,000 tw + 10,000 estab ip4 sockets. "manysock -6 10000 10000" - 10,000 tw + 10,000 estab ip6 sockets. Found at http://ilerner.3b1.org/manysock/manysock.c (3) Algorithmic analysis. Old algorithm. During 'cat </proc/net/tcp', tcp_seq_start() is called O(numsockets) times (4). On average, every call to tcp_seq_start() scans half the whole hashtable. Ouch. This is O(numsockets * hashsize). 95-99% of 'cat </proc/net/tcp' is spent in tcp_seq_start()->tcp_get_idx. This overhead is eliminated by new algorithm, which is O(numsockets + hashsize). New algorithm. New algorithms is O(numsockets + hashsize). We jump to the right hash bucket in tcp_seq_start(), without scanning half the hash. To jump right to the hash bucket corresponding to *pos in tcp_seq_start(), we reuse three pieces of state (st->num, st->bucket, st->sbucket) as follows: - we check that requested pos >= last seen pos (st->num), the typical case. - if so, we jump to bucket st->bucket - to arrive to the right item after beginning of st->bucket, we keep in st->sbucket the position corresponding to the beginning of bucket. (4) Explanation of O( numsockets * hashsize) of old algorithm. tcp_seq_start() is called once for every ~7 lines of netstat output if readsize is 1kb, or once for every ~28 lines if readsize >= 4kb. Since record length of /proc/net/tcp records is 150 bytes, formula for number of calls to tcp_seq_start() is (numsockets * 150 / min(4096,readsize)). Netstat uses 4kb readsize (newer versions), or 1kb (older versions). Note that speed of old algorithm does not improve above 4kb blocksize. Speed of the new algorithm does not depend on blocksize. Speed of the new algorithm does not perceptibly depend on hashsize (which depends on ramsize). Speed of old algorithm drops with bigger hashsize. (5) Reporting order. Reporting order is exactly same as before if hash does not change underfoot. When hash elements come and go during report, reporting order will be same as that of tcpdiag. Signed-off-by: Yakov Lerner <iler.ml@gmail.com> --- net/ipv4/tcp_ipv4.c | 26 ++++++++++++++++++++++++-- 1 files changed, 24 insertions(+), 2 deletions(-)