Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1471956 > unrolled thread

[RFC PATCH 0/6] Reduce cache miss for snmp_fold_field

Started byJia He <hejianet@gmail.com>
First post2016-08-29 18:10 +0200
Last post2016-08-29 18:10 +0200
Articles 4 — 1 participant

Back to article view | Back to linux.kernel

Contents

  [RFC PATCH 0/6] Reduce cache miss for snmp_fold_field Jia He <hejianet@gmail.com> - 2016-08-29 18:10 +0200
    [RFC PATCH 6/6] net: Suppress the "Comparison to NULL  Jia He <hejianet@gmail.com> - 2016-08-29 18:10 +0200
    [RFC PATCH 2/6] proc: Reduce cache miss in snmp6_seq_show Jia He <hejianet@gmail.com> - 2016-08-29 18:10 +0200
    [RFC PATCH 3/6] proc: Reduce cache miss in sctp_snmp_seq_show Jia He <hejianet@gmail.com> - 2016-08-29 18:10 +0200

#1471956 — [RFC PATCH 0/6] Reduce cache miss for snmp_fold_field

FromJia He <hejianet@gmail.com>
Date2016-08-29 18:10 +0200
Subject[RFC PATCH 0/6] Reduce cache miss for snmp_fold_field
Message-ID<sbxih-10O-9@gated-at.bofh.it>
In a PowerPc server with large cpu number(160), besides commit
a3a773726c9f ("net: Optimize snmp stat aggregation by walking all
the percpu data at once"), I watched several other snmp_fold_field
callsites which will cause high cache miss rate.

#My simple test case, which read from the procfs items endlessly:
/***********************************************************/
#include <stdlib.h>
#include <stdio.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#define LINELEN  2560
int main(int argc, char **argv)
{
        int i;
        int fd = -1 ;
        int rdsize = 0;
        char buf[LINELEN+1];

        buf[LINELEN] = 0;
        memset(buf,0,LINELEN);

        if(1 >= argc) {
                printf("file name empty\n");
                return -1;
        }

        fd = open(argv[1], O_RDWR, 0644);
        if(0 > fd){
                printf("open error\n");
                return -2;
        }

        for(i=0;i<0xffffffff;i++) {
                while(0 < (rdsize = read(fd,buf,LINELEN))){
                        //nothing here
                }

                lseek(fd, 0, SEEK_SET);
        }

        close(fd);
        return 0;
}
/**********************************************************/

#compile and run:
gcc test.c -o test

perf stat -d -e cache-misses ./test /proc/net/snmp
perf stat -d -e cache-misses ./test /proc/net/snmp6
perf stat -d -e cache-misses ./test /proc/net/netstat
perf stat -d -e cache-misses ./test /proc/net/sctp/snmp
perf stat -d -e cache-misses ./test /proc/net/xfrm_stat

#test results
I firstly test the correctness of data statistics. 

As for performance, before the patch set:
====================
 Performance counter stats for 'system wide':

         355911097      cache-misses                                                 [40.08%]
        2356829300      L1-dcache-loads                                              [60.04%]
         355642645      L1-dcache-load-misses     #   15.09% of all L1-dcache hits   [60.02%]
         346544541      LLC-loads                                                    [59.97%]
            389763      LLC-load-misses           #    0.11% of all LL-cache hits    [40.02%]

       6.245162638 seconds time elapsed

After the patch set:
===================
 Performance counter stats for 'system wide':

         194992476      cache-misses                                                 [40.03%]
        6718051877      L1-dcache-loads                                              [60.07%]
         194871921      L1-dcache-load-misses     #    2.90% of all L1-dcache hits   [60.11%]
         187632232      LLC-loads                                                    [60.04%]
            464466      LLC-load-misses           #    0.25% of all LL-cache hits    [39.89%]

       6.868422769 seconds time elapsed
The cache-miss rate can be reduced from 15% to 2.9%

Jia He (6):
  proc: Reduce cache miss in {snmp,netstat}_seq_show
  proc: Reduce cache miss in snmp6_seq_show
  proc: Reduce cache miss in sctp_snmp_seq_show
  proc: Reduce cache miss in xfrm_statistics_seq_show
  ipv6: Remove useless parameter in __snmp6_fill_statsdev
  net: Suppress the "Comparison to NULL could be written" warning

 net/ipv4/proc.c      | 110 ++++++++++++++++++++++++++++++---------------------
 net/ipv6/addrconf.c  |  12 +++---
 net/ipv6/proc.c      |  47 ++++++++++++++++------
 net/sctp/proc.c      |  15 +++++--
 net/xfrm/xfrm_proc.c |  15 +++++--
 5 files changed, 131 insertions(+), 68 deletions(-)

-- 
1.8.3.1

[toc] | [next] | [standalone]

#1471960 — [RFC PATCH 6/6] net: Suppress the "Comparison to NULL

FromJia He <hejianet@gmail.com>
Date2016-08-29 18:10 +0200
Subject[RFC PATCH 6/6] net: Suppress the "Comparison to NULL
Message-ID<sbxii-10O-43@gated-at.bofh.it>
In reply to#1471956
This is to suppress the checkpatch.pl warning "Comparison to NULL
could be written". No functional changes here. 

Signed-off-by: Jia He <hejianet@gmail.com>
---
 net/ipv4/proc.c      | 42 +++++++++++++++++++++---------------------
 net/sctp/proc.c      |  4 ++--
 net/xfrm/xfrm_proc.c |  4 ++--
 3 files changed, 25 insertions(+), 25 deletions(-)

diff --git a/net/ipv4/proc.c b/net/ipv4/proc.c
index ed09566..8cbf3e0 100644
--- a/net/ipv4/proc.c
+++ b/net/ipv4/proc.c
@@ -355,22 +355,22 @@ static void icmp_put(struct seq_file *seq)
 	atomic_long_t *ptr = net->mib.icmpmsg_statistics->mibs;
 
 	seq_puts(seq, "\nIcmp: InMsgs InErrors InCsumErrors");
-	for (i = 0; icmpmibmap[i].name != NULL; i++)
+	for (i = 0; icmpmibmap[i].name; i++)
 		seq_printf(seq, " In%s", icmpmibmap[i].name);
 	seq_puts(seq, " OutMsgs OutErrors");
-	for (i = 0; icmpmibmap[i].name != NULL; i++)
+	for (i = 0; icmpmibmap[i].name; i++)
 		seq_printf(seq, " Out%s", icmpmibmap[i].name);
 	seq_printf(seq, "\nIcmp: %lu %lu %lu",
 		snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_INMSGS),
 		snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_INERRORS),
 		snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_CSUMERRORS));
-	for (i = 0; icmpmibmap[i].name != NULL; i++)
+	for (i = 0; icmpmibmap[i].name; i++)
 		seq_printf(seq, " %lu",
 			   atomic_long_read(ptr + icmpmibmap[i].index));
 	seq_printf(seq, " %lu %lu",
 		snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_OUTMSGS),
 		snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_OUTERRORS));
-	for (i = 0; icmpmibmap[i].name != NULL; i++)
+	for (i = 0; icmpmibmap[i].name; i++)
 		seq_printf(seq, " %lu",
 			   atomic_long_read(ptr + (icmpmibmap[i].index | 0x100)));
 }
@@ -389,7 +389,7 @@ static int snmp_seq_show(struct seq_file *seq, void *v)
 	memset(buff64, 0, IPSTATS_MIB_MAX * sizeof(u64));
 
 	seq_puts(seq, "Ip: Forwarding DefaultTTL");
-	for (i = 0; snmp4_ipstats_list[i].name != NULL; i++)
+	for (i = 0; snmp4_ipstats_list[i].name; i++)
 		seq_printf(seq, " %s", snmp4_ipstats_list[i].name);
 
 	seq_printf(seq, "\nIp: %d %d",
@@ -399,30 +399,30 @@ static int snmp_seq_show(struct seq_file *seq, void *v)
 	BUILD_BUG_ON(offsetof(struct ipstats_mib, mibs) != 0);
 
 	for_each_possible_cpu(c) {
-		for (i = 0; snmp4_ipstats_list[i].name != NULL; i++)
+		for (i = 0; snmp4_ipstats_list[i].name; i++)
 			buff64[i] += snmp_get_cpu_field64(
 					net->mib.ip_statistics,
 					c, snmp4_ipstats_list[i].entry,
 					offsetof(struct ipstats_mib, syncp));
 	}
-	for (i = 0; snmp4_ipstats_list[i].name != NULL; i++)
+	for (i = 0; snmp4_ipstats_list[i].name; i++)
 		seq_printf(seq, " %llu", buff64[i]);
 
 	icmp_put(seq);	/* RFC 2011 compatibility */
 	icmpmsg_put(seq);
 
 	seq_puts(seq, "\nTcp:");
-	for (i = 0; snmp4_tcp_list[i].name != NULL; i++)
+	for (i = 0; snmp4_tcp_list[i].name; i++)
 		seq_printf(seq, " %s", snmp4_tcp_list[i].name);
 
 	seq_puts(seq, "\nTcp:");
 	for_each_possible_cpu(c) {
-		for (i = 0; snmp4_tcp_list[i].name != NULL; i++)
+		for (i = 0; snmp4_tcp_list[i].name; i++)
 			buff[i] += snmp_get_cpu_field(net->mib.tcp_statistics,
 						c, snmp4_tcp_list[i].entry);
 	}
 
-	for (i = 0; snmp4_tcp_list[i].name != NULL; i++) {
+	for (i = 0; snmp4_tcp_list[i].name; i++) {
 		/* MaxConn field is signed, RFC 2012 */
 		if (snmp4_tcp_list[i].entry == TCP_MIB_MAXCONN)
 			seq_printf(seq, " %ld", buff[i]);
@@ -433,23 +433,23 @@ static int snmp_seq_show(struct seq_file *seq, void *v)
 	memset(buff, 0, TCP_MIB_MAX * sizeof(unsigned long));
 
 	for_each_possible_cpu(c) {
-		for (i = 0; snmp4_udp_list[i].name != NULL; i++)
+		for (i = 0; snmp4_udp_list[i].name; i++)
 			buff[i] += snmp_get_cpu_field(net->mib.udp_statistics,
 						c, snmp4_udp_list[i].entry);
 	}
 	seq_puts(seq, "\nUdp:");
-	for (i = 0; snmp4_udp_list[i].name != NULL; i++)
+	for (i = 0; snmp4_udp_list[i].name; i++)
 		seq_printf(seq, " %s", snmp4_udp_list[i].name);
 	seq_puts(seq, "\nUdp:");
-	for (i = 0; snmp4_udp_list[i].name != NULL; i++)
+	for (i = 0; snmp4_udp_list[i].name; i++)
 		seq_printf(seq, " %lu", buff[i]);
 
 	/* the UDP and UDP-Lite MIBs are the same */
 	seq_puts(seq, "\nUdpLite:");
-	for (i = 0; snmp4_udp_list[i].name != NULL; i++)
+	for (i = 0; snmp4_udp_list[i].name; i++)
 		seq_printf(seq, " %s", snmp4_udp_list[i].name);
 	seq_puts(seq, "\nUdpLite:");
-	for (i = 0; snmp4_udp_list[i].name != NULL; i++)
+	for (i = 0; snmp4_udp_list[i].name; i++)
 		seq_printf(seq, " %lu", buff[i]);
 
 	seq_putc(seq, '\n');
@@ -485,29 +485,29 @@ static int netstat_seq_show(struct seq_file *seq, void *v)
 	memset(buff64, 0, sizeof(u64) * IPSTATS_MIB_MAX);
 
 	seq_puts(seq, "TcpExt:");
-	for (i = 0; snmp4_net_list[i].name != NULL; i++)
+	for (i = 0; snmp4_net_list[i].name; i++)
 		seq_printf(seq, " %s", snmp4_net_list[i].name);
 
 	seq_puts(seq, "\nTcpExt:");
 	for_each_possible_cpu(c)
-		for (i = 0; snmp4_net_list[i].name != NULL; i++)
+		for (i = 0; snmp4_net_list[i].name; i++)
 			buff[i] += snmp_get_cpu_field(net->mib.net_statistics,
 						c, snmp4_net_list[i].entry);
-	for (i = 0; snmp4_net_list[i].name != NULL; i++)
+	for (i = 0; snmp4_net_list[i].name; i++)
 		seq_printf(seq, " %lu", buff[i]);
 
 	seq_puts(seq, "\nIpExt:");
-	for (i = 0; snmp4_ipextstats_list[i].name != NULL; i++)
+	for (i = 0; snmp4_ipextstats_list[i].name; i++)
 		seq_printf(seq, " %s", snmp4_ipextstats_list[i].name);
 
 	seq_puts(seq, "\nIpExt:");
 	for_each_possible_cpu(c)
-		for (i = 0; snmp4_ipextstats_list[i].name != NULL; i++)
+		for (i = 0; snmp4_ipextstats_list[i].name; i++)
 			buff64[i] += snmp_get_cpu_field64(
 					net->mib.ip_statistics,
 					c, snmp4_ipextstats_list[i].entry,
 					offsetof(struct ipstats_mib, syncp));
-	for (i = 0; snmp4_ipextstats_list[i].name != NULL; i++)
+	for (i = 0; snmp4_ipextstats_list[i].name; i++)
 		seq_printf(seq, " %llu", buff64[i]);
 
 	seq_putc(seq, '\n');
diff --git a/net/sctp/proc.c b/net/sctp/proc.c
index 085fb95..816a5e8 100644
--- a/net/sctp/proc.c
+++ b/net/sctp/proc.c
@@ -80,11 +80,11 @@ static int sctp_snmp_seq_show(struct seq_file *seq, void *v)
 	memset(buff, 0, sizeof(unsigned long) * SCTP_MIB_MAX);
 
 	for_each_possible_cpu(c)
-		for (i = 0; sctp_snmp_list[i].name != NULL; i++)
+		for (i = 0; sctp_snmp_list[i].name; i++)
 			buff[i] += snmp_get_cpu_field(
 						net->sctp.sctp_statistics,
 						c, sctp_snmp_list[i].entry);
-	for (i = 0; sctp_snmp_list[i].name != NULL; i++)
+	for (i = 0; sctp_snmp_list[i].name; i++)
 		seq_printf(seq, "%-32s\t%ld\n", sctp_snmp_list[i].name,
 						buff[i]);
 
diff --git a/net/xfrm/xfrm_proc.c b/net/xfrm/xfrm_proc.c
index c9df546..2f1da2d 100644
--- a/net/xfrm/xfrm_proc.c
+++ b/net/xfrm/xfrm_proc.c
@@ -57,11 +57,11 @@ static int xfrm_statistics_seq_show(struct seq_file *seq, void *v)
 	memset(buff, 0, sizeof(unsigned long) * LINUX_MIB_XFRMMAX);
 
 	for_each_possible_cpu(c)
-		for (i = 0; xfrm_mib_list[i].name != NULL; i++)
+		for (i = 0; xfrm_mib_list[i].name; i++)
 			buff[i] += snmp_get_cpu_field(
 						net->mib.xfrm_statistics,
 						c, xfrm_mib_list[i].entry);
-	for (i = 0; xfrm_mib_list[i].name != NULL; i++)
+	for (i = 0; xfrm_mib_list[i].name; i++)
 		seq_printf(seq, "%-24s\t%lu\n", xfrm_mib_list[i].name,
 						buff[i]);
 
-- 
1.8.3.1

[toc] | [prev] | [next] | [standalone]

#1471962 — [RFC PATCH 2/6] proc: Reduce cache miss in snmp6_seq_show

FromJia He <hejianet@gmail.com>
Date2016-08-29 18:10 +0200
Subject[RFC PATCH 2/6] proc: Reduce cache miss in snmp6_seq_show
Message-ID<sbxii-10O-51@gated-at.bofh.it>
In reply to#1471956
This patch exchanges the two loop for collecting the percpu
statistics data. This can reduce cache misses by going through
all the items of each cpu sequentially.

Signed-off-by: Jia He <hejianet@gmail.com>
---
 net/ipv6/proc.c | 47 ++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 36 insertions(+), 11 deletions(-)

diff --git a/net/ipv6/proc.c b/net/ipv6/proc.c
index 679253d0..c834646 100644
--- a/net/ipv6/proc.c
+++ b/net/ipv6/proc.c
@@ -30,6 +30,13 @@
 #include <net/transp_v6.h>
 #include <net/ipv6.h>
 
+#define MAX(a, b) ((u32)(a) >= (u32)(b) ? (a) : (b))
+
+#define MAX4(a, b, c, d) \
+	MAX(MAX(a, b), MAX(c, d))
+#define SNMP_MIB_MAX MAX4(UDP_MIB_MAX, TCP_MIB_MAX, \
+			IPSTATS_MIB_MAX, ICMP_MIB_MAX)
+
 static int sockstat6_seq_show(struct seq_file *seq, void *v)
 {
 	struct net *net = seq->private;
@@ -191,25 +198,43 @@ static void snmp6_seq_show_item(struct seq_file *seq, void __percpu *pcpumib,
 				atomic_long_t *smib,
 				const struct snmp_mib *itemlist)
 {
-	int i;
-	unsigned long val;
-
-	for (i = 0; itemlist[i].name; i++) {
-		val = pcpumib ?
-			snmp_fold_field(pcpumib, itemlist[i].entry) :
-			atomic_long_read(smib + itemlist[i].entry);
-		seq_printf(seq, "%-32s\t%lu\n", itemlist[i].name, val);
+	int i, c;
+	unsigned long buff[SNMP_MIB_MAX];
+
+	memset(buff, 0, sizeof(unsigned long) * SNMP_MIB_MAX);
+
+	if (pcpumib) {
+		for_each_possible_cpu(c)
+			for (i = 0; itemlist[i].name; i++)
+				buff[i] += snmp_get_cpu_field(pcpumib, c,
+							itemlist[i].entry);
+		for (i = 0; itemlist[i].name; i++)
+			seq_printf(seq, "%-32s\t%lu\n",
+				   itemlist[i].name, buff[i]);
+	} else {
+		for (i = 0; itemlist[i].name; i++)
+			seq_printf(seq, "%-32s\t%lu\n", itemlist[i].name,
+				   atomic_long_read(smib + itemlist[i].entry));
 	}
 }
 
 static void snmp6_seq_show_item64(struct seq_file *seq, void __percpu *mib,
 				  const struct snmp_mib *itemlist, size_t syncpoff)
 {
-	int i;
+	int i, c;
+	u64 buff[SNMP_MIB_MAX];
+
+	memset(buff, 0, sizeof(unsigned long) * SNMP_MIB_MAX);
 
+	for_each_possible_cpu(c) {
+		for (i = 0; itemlist[i].name; i++) {
+			buff[i] += snmp_get_cpu_field64(mib, c,
+							itemlist[i].entry,
+							syncpoff);
+		}
+	}
 	for (i = 0; itemlist[i].name; i++)
-		seq_printf(seq, "%-32s\t%llu\n", itemlist[i].name,
-			   snmp_fold_field64(mib, itemlist[i].entry, syncpoff));
+		seq_printf(seq, "%-32s\t%llu\n", itemlist[i].name, buff[i]);
 }
 
 static int snmp6_seq_show(struct seq_file *seq, void *v)
-- 
1.8.3.1

[toc] | [prev] | [next] | [standalone]

#1471963 — [RFC PATCH 3/6] proc: Reduce cache miss in sctp_snmp_seq_show

FromJia He <hejianet@gmail.com>
Date2016-08-29 18:10 +0200
Subject[RFC PATCH 3/6] proc: Reduce cache miss in sctp_snmp_seq_show
Message-ID<sbxii-10O-49@gated-at.bofh.it>
In reply to#1471956
This patch exchanges the two loop for collecting the percpu
statistics data. This can reduce cache misses by going through
all the items of each cpu sequentially.

Signed-off-by: Jia He <hejianet@gmail.com>
---
 net/sctp/proc.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/net/sctp/proc.c b/net/sctp/proc.c
index ef8ba77..085fb95 100644
--- a/net/sctp/proc.c
+++ b/net/sctp/proc.c
@@ -74,12 +74,19 @@ static const struct snmp_mib sctp_snmp_list[] = {
 static int sctp_snmp_seq_show(struct seq_file *seq, void *v)
 {
 	struct net *net = seq->private;
-	int i;
+	int i, c;
+	unsigned long buff[SCTP_MIB_MAX];
 
+	memset(buff, 0, sizeof(unsigned long) * SCTP_MIB_MAX);
+
+	for_each_possible_cpu(c)
+		for (i = 0; sctp_snmp_list[i].name != NULL; i++)
+			buff[i] += snmp_get_cpu_field(
+						net->sctp.sctp_statistics,
+						c, sctp_snmp_list[i].entry);
 	for (i = 0; sctp_snmp_list[i].name != NULL; i++)
 		seq_printf(seq, "%-32s\t%ld\n", sctp_snmp_list[i].name,
-			   snmp_fold_field(net->sctp.sctp_statistics,
-				      sctp_snmp_list[i].entry));
+						buff[i]);
 
 	return 0;
 }
-- 
1.8.3.1

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web