1

I am testing the performance of syscall on different kernel(The hardware are same):

The test code

#define _GNU_SOURCE
#include <unistd.h>
#include <sys/syscall.h>
#include <sys/types.h>

int main(int argc, char *argv[])
{
       int i;
       pid_t tid;
       unsigned long cnt = 1000000;

       for(i = 0; i < cnt; i++)
       {
               tid = syscall(SYS_gettid);
       }
       return 0;
}

The result of kernel 2.6:

processor       : 3
vendor_id       : GenuineIntel
cpu family      : 6
model           : 55
model name      : Intel(R) Celeron(R) CPU  J1900  @ 1.99GHz
stepping        : 9
cpu MHz         : 2000.029
cache size      : 1024 KB
physical id     : 0
siblings        : 4
core id         : 3
cpu cores       : 4
apicid          : 6
initial apicid  : 6
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 11
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx rdtscp lm constant_tsc arch_perfmon pebs bts xtopology nonstop_tsc aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm sse4_1 sse4_2 movbe popcnt lahf_lm 3dnowprefetch arat tpr_shadow vnmi flexpriority ept vpid
bogomips        : 3999.80
clflush size    : 64
cache_alignment : 64
address sizes   : 36 bits physical, 48 bits virtual
power management:

[HFOS] $ uname -a
Linux HFOS 2.6.32.10 #1 SMP Fri Sep 9 16:11:47 CST 2016 i686 pentium3 i386 GNU/Linux

[HFOS] $ strace -c ./sc
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
99.90    0.023803           0   1000000           gettid
 0.10    0.000023           1        40        39 open
 0.00    0.000000           0         1           read
 0.00    0.000000           0         1           close
 0.00    0.000000           0         1           execve
 0.00    0.000000           0         1         1 access
 0.00    0.000000           0         1           brk
 0.00    0.000000           0         5           mmap2
 0.00    0.000000           0        39        35 stat64
 0.00    0.000000           0         1           fstat64
 0.00    0.000000           0         1           set_thread_area
------ ----------- ----------- --------- --------- ----------------
100.00    0.023826               1000091        75 total

The result of kernel 5.2:

processor       : 3
vendor_id       : GenuineIntel
cpu family      : 6
model           : 55
model name      : Intel(R) Celeron(R) CPU  J1900  @ 1.99GHz
stepping        : 9
microcode       : 0x90a
cpu MHz         : 1332.848
cache size      : 1024 KB
physical id     : 0
siblings        : 4
core id         : 3
cpu cores       : 4
apicid          : 6
initial apicid  : 6
fpu             : yes
fpu_exception   : yes
cpuid level     : 11
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology tsc_reliable nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm sse4_1 sse4_2 movbe popcnt tsc_deadline_timer rdrand lahf_lm 3dnowprefetch epb pti ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid tsc_adjust smep erms dtherm arat
bugs            : cpu_meltdown spectre_v1 spectre_v2 mds msbds_only
bogomips        : 3998.40
clflush size    : 64
cache_alignment : 64
address sizes   : 36 bits physical, 48 bits virtual
power management:

localhost:ipc # uname -a
Linux localhost 5.2.8 #2 SMP Wed May 6 12:51:13 CST 2020 x86_64 GNU/Linux

strace -c ./sc
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
99.99    7.964185           7   1000000           gettid
 0.00    0.000384         384         1           execve
 0.00    0.000152          21         7           mmap
 0.00    0.000084          21         4           mprotect
 0.00    0.000045          22         2           openat
 0.00    0.000031          31         1           munmap
 0.00    0.000027          13         2           fstat
 0.00    0.000021          10         2           close
 0.00    0.000020          20         1         1 access
 0.00    0.000016          16         1           read
 0.00    0.000011          11         1           brk
 0.00    0.000010          10         1           arch_prctl
------ ----------- ----------- --------- --------- ----------------
100.00    7.964986           7   1000023         1 total

I am confused why it is so slow on new kernel. Please help me. Thanks very much.

After setting [mitigations=off] the syscall performance are almost same(The excute time are almost same, but strace time are different).

localhost:~ # dmesg | grep iso
[    0.006693] Kernel/User page tables isolation: disabled on command line.

The result is shown bellow.

Kernel 2.6:

[HFOS] $ uname -a
Linux HFOS 2.6.32.10 #1 SMP Fri Sep 9 16:11:47 CST 2016 i686 pentium3 i386 GNU/Linux
[HFOS] $ cat sc.c
#define _GNU_SOURCE
#include <unistd.h>
#include <sys/syscall.h>
#include <sys/types.h>

int main(int argc, char *argv[])
{
    int i;
    pid_t tid;
    unsigned long cnt = 100000000;

    for(i = 0; i < cnt; i++)
    {
        tid = syscall(SYS_gettid);
    }
    return 0;
}

[HFOS] $ time ./sc

real    0m16.736s
user    0m5.529s
sys     0m11.204s

[HFOS] $ time strace -c ./sc
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
100.00    0.275026           0  10000000           gettid
  0.00    0.000000           0         1           read
  0.00    0.000000           0        40        39 open
  0.00    0.000000           0         1           close
  0.00    0.000000           0         1           execve
  0.00    0.000000           0         1         1 access
  0.00    0.000000           0         1           brk
  0.00    0.000000           0         5           mmap2
  0.00    0.000000           0        39        35 stat64
  0.00    0.000000           0         1           fstat64
  0.00    0.000000           0         1           set_thread_area
------ ----------- ----------- --------- --------- ----------------
100.00    0.275026              10000091        75 total

real    2m57.054s
user    0m28.704s
sys     2m27.259s



Kernel 5.2:

localhost:test # uname -a
Linux localhost 5.2.8 #2 SMP Thu May 14 02:46:43 CST 2020 x86_64 GNU/Linux
localhost:test # cat sc.c
#define _GNU_SOURCE
#include <unistd.h>
#include <sys/syscall.h>
#include <sys/types.h>

int main(int argc, char *argv[])
{
    int i;
    pid_t tid;
    unsigned long cnt = 100000000;

    for(i = 0; i < cnt; i++)
    {
        tid = syscall(SYS_gettid);
    }
    return 0;
}

localhost:test # time ./sc

real    0m19.043s
user    0m8.501s
sys     0m10.532s

localhost:test # time strace -c ./sc
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
100.00   77.250398           7  10000000           gettid
  0.00    0.000405         405         1           execve
  0.00    0.000159          22         7           mmap
  0.00    0.000088          22         4           mprotect
  0.00    0.000048          24         2           openat
  0.00    0.000031          31         1           munmap
  0.00    0.000028          14         2           fstat
  0.00    0.000024          12         2           close
  0.00    0.000021          21         1         1 access
  0.00    0.000016          16         1           read
  0.00    0.000013          13         1           brk
  0.00    0.000012          12         1           arch_prctl
------ ----------- ----------- --------- --------- ----------------
100.00   77.251243           7  10000023         1 total

real    6m7.443s
user    0m55.590s
sys     6m23.482s


But the UNIX domain socket performance has no changed. The test code are shown below.

svr.c:

#include <unistd.h>
#include <errno.h>
#include <stdio.h>
#include <ctype.h>
#include <sys/un.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <arpa/inet.h>

#define TEST_SOCK_FILE  "/tmp/test.sock"

int main(int argc, char **argv)
{
    int fd;
    int cfd;
    int r;
    int cnt = 0;
    socklen_t  sklen;
    struct sockaddr caddr;
    char rbuf[1024];
    char sbuf[512];
    struct sockaddr_un  svraddr;

    unlink(TEST_SOCK_FILE);
    svraddr.sun_family = AF_UNIX;
    snprintf(svraddr.sun_path, sizeof(svraddr.sun_path), TEST_SOCK_FILE);

    fd = socket(AF_UNIX, SOCK_STREAM, 0);
    if(fd < 0)
    {
        printf("Create socket failed : %s\n", strerror(errno));
        return -1;
    }

    if(bind(fd, (struct sockaddr *)&svraddr, sizeof(svraddr)) < 0)
    {
        printf("Bind socket failed : %s\n", strerror(errno));
        close(fd);
        return -1;
    }

    if(listen(fd, 10) < 0)
    {
        printf("Listen socket failed : %s\n", strerror(errno));
        close(fd);
        return -1;
    }

    while(1)
    {
        sklen = sizeof(caddr);
        memset(&caddr, 0, sizeof(caddr));
        cfd = accept(fd, &caddr, &sklen);
        if(cfd < 0)
        {
            printf("Accept failed : %s", strerror(errno));
            return -1;
        }

        cnt = 0;
        while(1)
        {
            r = read(cfd, rbuf, sizeof(rbuf));
            if(r <= 0)
            {
                printf("recv failed : %s\n", strerror(errno));
                break;
            }
            if(rbuf[0] == 0x22)
            {
                break;
            }
            r = write(cfd, sbuf, sizeof(sbuf));
            if(r <= 0)
            {
                printf("send failed : %s\n", strerror(errno));
                break;
            }
            cnt++;
        }

        printf("Recv packet : %d\n", cnt);
        close(cfd);
    }
    close(fd);
    return 0;
}

cli.c:

#include <unistd.h>
#include <stdlib.h>
#include <stdio.h>
#include <ctype.h>
#include <time.h>
#include <errno.h>
#include <sys/un.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <arpa/inet.h>


int main(int argc, char **argv)
{
    int fd;
    int r;
    int i;
    int cnt = 1000000;
    struct sockaddr_un unaddr;
    time_t ts, te, tu;
    char sbuf[1024];
    char rbuf[512];

    unaddr.sun_family = AF_UNIX;
    strcpy(unaddr.sun_path, "/tmp/test.sock");
    fd = socket(SOCK_STREAM, SOCK_STREAM, 0);
    if(fd < 0)
    {
        printf("Create socket failed : %s\n", strerror(errno));
        return -1;
    }

    r = connect(fd, (struct sockaddr *)&unaddr, sizeof(unaddr));
    if(r < 0)
    {
        printf("Connect failed : %s\n", strerror(errno));
        close(fd);
        return -1;
    }

    ts = time(NULL);
    for(i = 0; i < cnt; i++)
    {
        sbuf[0] = 0x11;
        r = write(fd, sbuf, sizeof(sbuf));
        if(r <= 0)
        {
            printf("Send failed : %s\n", strerror(errno));
            break;
        }
        r = read(fd, rbuf, sizeof(rbuf));
        if(r <= 0)
        {
            printf("Recv failed : %s\n", strerror(errno));
            break;
        }
    }

    sbuf[0] = 0x22;
    write(fd, sbuf, sizeof(sbuf));

    te = time(NULL);
    tu = te > ts ? (te - ts) : 1;
    printf("PPS(%d) : %d packet used %lu seconds\n", cnt / tu, cnt, tu);

    close(fd);
    return 0;
}

The results are show bellow:

localhost:test # uname -a
Linux localhost 5.2.8 #2 SMP Thu May 14 02:46:43 CST 2020 x86_64 GNU/Linux
localhost:test # ./svr &
[1] 955
localhost:test # ./cli
PPS(34482) : 1000000 packet used 29 seconds
Recv packet : 1000000
[HFOS] $ uname -a
Linux HFOS 2.6.32.10 #1 SMP Fri Sep 9 16:11:47 CST 2016 i686 pentium3 i386 GNU/Linux
[HFOS] $ ./svr &
[1] 32624
[HFOS] $ ./cli
Recv packet : 1000000
PPS(71428) : 1000000 packet used 14 seconds

4
  • Have you installed the meltdown/spectre patches on the newer system? Context switching was specifically impacted by those security patches.
    – kemotep
    May 13, 2020 at 14:06
  • Yes, I installed those patches. I don't know these security pathes impact the performance so much. I will compare it with system without those patches. Thanks very much.
    – afeiraul
    May 13, 2020 at 14:58
  • its certainly not the only reason but a part of the problem may lay in the fact that the reported cpu clock is lower in the new kernel...
    – der bender
    May 13, 2020 at 17:59
  • The hardware and CPUs are all same. I think the reason of different CPU reported is Dynamic Power Management. I saw the CPU was same by /proc/cpuinfo when CPU was busy.
    – afeiraul
    May 14, 2020 at 4:37

1 Answer 1

6

Most of the lower syscall performance on your 5.2 kernel compared to the 2.6.32 kernel is probably attributable to kernel page-table isolation and other security-related changes. KPTI involves using different page tables when running in user space and running in the kernel. Thus every system call changes the page tables, twice, with cascading consequences such as TLB flushes (on older hardware).

An interesting paper tracking the performance changes in the Linux kernel was published last year; it is described in some detail in this blog post, and the paper itself is available on ACM DL (and publicly accessible until the end of June).

3
  • Thanks very much. I am writing a UNIX Domain Socket server and clients. The client send 1024 bytes data to server, and the server send 512 bytes to client. I am surprised that the performance on kernel 5.2 is only 50% of kernel 2.6. I am very sad to found the result. I will research the information you provided and try to find the resolution.
    – afeiraul
    May 13, 2020 at 14:59
  • See this answer for details of the kernel parameters you can use to toggle the security mitigations; that should help you narrow down the cause of the slowdown (assuming it is related to those mitigations). May 13, 2020 at 15:01
  • Thanks for helpful information. I tested to set mitigations=off and the syscall performance were almost same on 2 kernels(5.26 Million/S on 5.2 and 5.88Million/S on 2.6). But UNIX domain socket performance has no changed. I will post the test code below. Waiting for your helpful advice.
    – afeiraul
    May 14, 2020 at 8:21

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .