Hi, Vladimir Homutov:
The below is our WRK test result output with BPF enable
112 threads and 10000 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 608.23ms 820.71ms 10.00s 87.48%
Connect 16.52ms 54.53ms 1.99s 94.73%
Delay 153.13ms 182.17ms 2.00s 90.74%
Req/Sec 244.79 142.32 1.99k 68.40%
Latency Distribution
50.00% 293.50ms
75.00% 778.33ms
90.00% 1.61s
99.00% 3.71s
99.90% 7.03s
99.99% 8.94s
Connect Distribution
50.00% 1.93ms
75.00% 2.85ms
90.00% 55.76ms
99.00% 229.19ms
99.90% 656.79ms
99.99% 1.43s
Delay Distribution
50.00% 110.96ms
75.00% 193.67ms
90.00% 321.77ms
99.00% 959.27ms
99.90% 1.57s
99.99% 1.91s
Compared with no BPF but enable reuseport as below
112 threads and 10000 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 680.50ms 943.69ms 10.00s 87.18%
Connect 58.44ms 238.08ms 2.00s 94.58%
Delay 158.84ms 256.28ms 2.00s 90.92%
Req/Sec 244.51 151.00 1.41k 69.67%
Latency Distribution
50.00% 317.61ms
75.00% 913.52ms
90.00% 1.90s
99.00% 4.30s
99.90% 6.52s
99.99% 8.80s
Connect Distribution
50.00% 1.88ms
75.00% 2.21ms
90.00% 55.94ms
99.00% 1.45s
99.90% 1.95s
99.99% 2.00s
Delay Distribution
50.00% 73.01ms
75.00% 190.40ms
90.00% 387.01ms
99.00% 1.34s
99.90% 1.86s
99.99% 1.99s
From the above results, there shows almost 20% percent latency reduction. P99 latency of BPF is 3.71s , but without BPF is 4.3s.
For question2 why not bind queues to cores, as what we think ,bind to thread is different from bind to core. Bind to threads is user space work to distribute the load, bind to core is NIC hardware level.
We do have ADQ for NIC, that will do a lot bind to cores things
Thanks
LQ
-----Original Message-----
From: nginx-devel <nginx-devel-bounces@nginx.org> On Behalf Of Vladimir Homutov
Sent: Thursday, September 10, 2020 6:12 PM
To: nginx-devel@nginx.org
Subject: Re: [PATCH] Use BPF to distribute packet to different work thread.
On Thu, Sep 10, 2020 at 05:22:12AM +0000, Liu, Qiao wrote:
> # HG changeset patch
> # User Liu Qiao <qiao.liu@intel.com>
> # Date 1599735293 14400
> # Thu Sep 10 06:54:53 2020 -0400
> # Node ID f79d524a2cc0093c53490f947564e42371cf944f
> # Parent da5e3f5b16733167142b599b6af3ce9469a07d52
> Use BPF to distribute packet to different work thread.
> Use Berkeley Packet Filter to get packet queue_mapping number, and use
> this queue_mapping number to distribute the packet to different work
> thread, this will improve CPU utilization and http latency.
> Author: Samudrala, Sridhar <sridhar.samudrala@intel.com>
Do you have any real measurements with proper testing methodology?
Also, why not bind NIC queues to cores system-wide?
>
> diff -r da5e3f5b1673 -r f79d524a2cc0 auto/os/linux
> --- a/auto/os/linux Wed Sep 02 23:13:36 2020 +0300
> +++ b/auto/os/linux Thu Sep 10 06:54:53 2020 -0400
> @@ -32,6 +32,10 @@
> have=NGX_HAVE_POSIX_FADVISE . auto/nohave fi
> +if [ $version -lt 263680 ]; then
> + have=NGX_HAVE_REUSEPORT_CBPF . auto/nohave fi
> +
> # epoll, EPOLLET version
> ngx_feature="epoll"
> diff -r da5e3f5b1673 -r f79d524a2cc0 auto/unix
> --- a/auto/unix Wed Sep 02 23:13:36 2020 +0300
> +++ b/auto/unix Thu Sep 10 06:54:53 2020 -0400
> @@ -331,6 +331,17 @@
> ngx_feature_test="setsockopt(0, SOL_SOCKET, SO_REUSEPORT, NULL, 0)"
> . auto/feature
> +ngx_feature="SO_REUSEPORT_CBPF"
> +ngx_feature_name="NGX_HAVE_REUSEPORT_CBPF"
> +ngx_feature_run=no
> +ngx_feature_incs="#include <sys/socket.h>
> + #include <linux/filter.h>
> + #include <error.h>"
> +ngx_feature_path=
> +ngx_feature_libs=
> +ngx_feature_test="setsockopt(0, SOL_SOCKET, SO_ATTACH_REUSEPORT_CBPF, NULL, 0)"
> +. auto/feature
> +
> ngx_feature="SO_ACCEPTFILTER"
> ngx_feature_name="NGX_HAVE_DEFERRED_ACCEPT"
> diff -r da5e3f5b1673 -r f79d524a2cc0 src/core/ngx_connection.c
> --- a/src/core/ngx_connection.c Wed Sep 02 23:13:36 2020 +0300
> +++ b/src/core/ngx_connection.c Thu Sep 10 06:54:53 2020 -0400
> @@ -8,7 +8,10 @@
> #include <ngx_config.h>
> #include <ngx_core.h>
> #include <ngx_event.h>
> -
> +#if (NGX_HAVE_REUSEPORT_CBPF)
> +#include <linux/filter.h>
> +#include <error.h>
> +#endif
> ngx_os_io_t ngx_io;
> @@ -708,6 +711,35 @@
> return NGX_OK;
> }
> +#if(NGX_HAVE_REUSEPORT)
> +#if(NGX_HAVE_REUSEPORT_CBPF)
> +#ifndef ARRAY_SIZE
> +#define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0])) #endif
> +
> +static void attach_bpf(int fd, uint16_t n) {
> + struct sock_filter code[] = {
> + /* A = skb->queue_mapping */
> + { BPF_LD | BPF_W | BPF_ABS, 0, 0, SKF_AD_OFF + SKF_AD_QUEUE },
> + /* A = A % n */
> + { BPF_ALU | BPF_MOD, 0, 0, n },
> + /* return A */
> + { BPF_RET | BPF_A, 0, 0, 0 },
> + };
> + struct sock_fprog p = {
> + .len = ARRAY_SIZE(code),
> + .filter = code,
> + };
> +
> + if (setsockopt(fd, SOL_SOCKET, SO_ATTACH_REUSEPORT_CBPF, &p, sizeof(p)))
> + printf("failed to set SO_ATTACH_REUSEPORT_CBPF");
> + else
> + printf("bpf prog attached to fd:%d\n",
> +fd); } #endif #endif
> +
> void
> ngx_configure_listening_sockets(ngx_cycle_t *cycle) @@ -719,6 +751,11
> @@ #if (NGX_HAVE_DEFERRED_ACCEPT && defined SO_ACCEPTFILTER)
> struct accept_filter_arg af;
> #endif
> +#if (NGX_HAVE_REUSEPORT)
> +#if (NGX_HAVE_REUSEPORT_CBPF)
> + ngx_core_conf_t* ccf ;
> +#endif
> +#endif
> ls = cycle->listening.elts;
> for (i = 0; i < cycle->listening.nelts; i++) { @@ -1011,6
> +1048,16 @@
> }
> #endif
> +#if (NGX_HAVE_REUSEPORT)
> +#if (NGX_HAVE_REUSEPORT_CBPF)
> + if(ls[i].reuseport)
> + {
> + ccf = (ngx_core_conf_t *) ngx_get_conf(cycle->conf_ctx,ngx_core_module);
> + if(ccf)
> + attach_bpf(ls[i].fd, ccf->worker_processes);
> + }
> +#endif
> +#endif
> }
> return;
>
> _______________________________________________
> nginx-devel mailing list
> nginx-devel@nginx.org
> http://mailman.nginx.org/mailman/listinfo/nginx-devel
_______________________________________________
nginx-devel mailing list
nginx-devel@nginx.org
http://mailman.nginx.org/mailman/listinfo/nginx-devel
_______________________________________________
nginx-devel mailing list
nginx-devel@nginx.org
http://mailman.nginx.org/mailman/listinfo/nginx-devel
