Skip to main content
Version: 5.0 (Stable) ✅

Performance

There is a set of tools for performance improvement and detecting memory leaking.

Note: All tools will hurts performance more or less, so never enable these tools unless you need to fix memory issue.

RTC

RTC is delivering over UDP, so the first and most important configuration is for kernel network:

# Query the kernel configuration
sysctl net.core.rmem_max
sysctl net.core.rmem_default
sysctl net.core.wmem_max
sysctl net.core.wmem_default

# Set the UDP buffer to 16MB
sysctl net.core.rmem_max=16777216
sysctl net.core.rmem_default=16777216
sysctl net.core.wmem_max=16777216
sysctl net.core.wmem_default=16777216

Note: For Docker, it read the configuration from host, so you only need to setup the host machine.

Note:If need to set these configurations in docker, you must run with --network=host.

Or, you could also modify the file /etc/sysctl.conf to enalbe if when reboot:

# vi /etc/sysctl.conf
# For RTC
net.core.rmem_max=16777216
net.core.rmem_default=16777216
net.core.wmem_max=16777216
net.core.wmem_default=16777216

Query the network statistics and UDP packets dropping:

netstat -suna
netstat -suna && sleep 30 && netstat -suna

For Example:

  • 224911319 packets received The total received UDP packets.
  • 65731106 receive buffer errors The total dropped UDP packets before receiving
  • 123534411 packets sent The total sent UDP packets.
  • 0 send buffer errors The total dropped UDP packets before sending.

Note: SRS also prints about the packets dropped in application level, for example loss=(r:49,s:0) which means dropped 49 packets before receiving.

Note:Please note that you must run the command in docker container, not on host machine.

The length of UDP queue:

netstat -lpun

For example:

  • Recv-Q 427008 Established: The count of bytes not copied by the user program connected to this socket.
  • Send-Q 0 Established: The count of bytes not acknowledged by the remote host.

Other useful parameters of netstat:

  • --udp|-u Filter by UDP protocol.
  • --numeric|-n Show numerical addresses instead of trying to determine symbolic host, port or user names.
  • --statistics|-s Show statistics.
  • --all|-a Show both listening and non-listening sockets. With the --interfaces option, show interfaces that are not up.
  • --listening|-l Show only listening sockets. (These are omitted by default.)
  • --program|-p Show the PID and name of the program to which each socket belongs.

PERF

PERF is Performance analysis tools for Linux.

Show performance bottleneck of SRS:

perf top -p $(pidof srs)

To record the data:

perf record -p $(pidof srs)

# Press CTRL+C after about 30s.

perf report

Show stack or backtrace:

perf record -a --call-graph fp -p $(pidof srs)
perf report --call-graph --stdio

Note: Record to file by perf report --call-graph --stdio >t.txt

Remark: The stack(-g) does not work for SRS(ST), because ST modifies the SP.

ASAN(Google Address Sanitizer)

SRS5+ supports ASAN by default.

If you want to disable it, please check bellow configure options:

./configure -h |grep asan
  --sanitizer=on|off        Whether build SRS with address sanitizer(asan). Default: on
  --sanitizer-static=on|off Whether build SRS with static libasan(asan). Default: off
  --sanitizer-log=on|off    Whether hijack the log for libasan(asan). Default: off

Highly recommend to enable ASAN because it works great.

GPROF

GPROF is a GNU tool, see SRS GPROF and GNU GPROF.

Usage:

# Build SRS with GPROF
./configure --gprof=on && make

# Start SRS with GPROF
./objs/srs -c conf/console.conf

# Or CTRL+C to stop GPROF
killall -2 srs

# To analysis result.
gprof -b ./objs/srs gmon.out

GPERF

GPERF is google tcmalloc, please see GPERF

GPERF: GCP

GCP is for CPU performance analysis, see GCP.

Usage:

# Build SRS with GCP
./configure --gperf=on --gcp=on && make

# Start SRS with GCP
./objs/srs -c conf/console.conf

# Or CTRL+C to stop GCP
killall -2 srs

# To analysis cpu profile
./objs/pprof --text objs/srs gperf.srs.gcp*

Note: For more details, please read cpu-profiler.

Install tool for graph:

yum install -y graphviz

Output svg graph to open by Chrome:

./objs/pprof --svg ./objs/srs gperf.srs.gcp >t.svg

GPERF: GMD

GMD is for memory corrupt detecting, see GMD.

Usage:

# Build SRS with GMD.
./configure --gperf=on --gmd=on && make

# Start SRS with GMD.
env TCMALLOC_PAGE_FENCE=1 ./objs/srs -c conf/console.conf

Note: For more details, please read heap-defense.

Note: Need link with libtcmalloc_debug.a and enable env TCMALLOC_PAGE_FENCE.

GPERF: GMC

GMC is for memory leaking, see GMC.

Usage:

# Build SRS with GMC
./configure --gperf=on --gmc=on && make

# Start SRS with GMC
env PPROF_PATH=./objs/pprof HEAPCHECK=normal ./objs/srs -c conf/console.conf 2>gmc.log 

# Or CTRL+C to stop gmc
killall -2 srs

# To analysis memory leak
cat gmc.log

Note: For more details, please read heap-checker.

GPERF: GMP

GMD is for memory performance, see GMP.

Usage:

# Build SRS with GMP
./configure --gperf=on --gmp=on && make

# Start SRS with GMP
./objs/srs -c conf/console.conf

# Or CTRL+C to stop gmp
killall -2 srs 

# To analysis memory profile
./objs/pprof --text objs/srs gperf.srs.gmp*

Note: For more details, please read heap-profiler.

VALGRIND

SRS3+ also supports valgrind.

valgrind --leak-check=full ./objs/srs -c conf/console.conf

Remark: For ST to support valgrind, see state-threads and ST#2.

Syscall

Please use strace -c -p PID for syscal performance issue.

OSX

For macOS, please use Instruments

instruments -l 30000 -t Time\ Profiler -p 72030

Remark: You can also click Sample button in Active Monitor.

Multiple Process and Softirq

You can run softirq(Kernel Network Transmission) on CPU0, so run SRS on other CPUs:

taskset -p 0xfe $(pidof srs)

Or run SRS on CPU1:

taskset -pc 1 $(pidof srs)

Then you can run top and press 1 to see each CPU statistics:

top # Press 1
#%Cpu0  :  1.8 us,  1.1 sy,  0.0 ni, 90.8 id,  0.0 wa,  0.0 hi,  6.2 si,  0.0 st
#%Cpu1  : 67.6 us, 17.6 sy,  0.0 ni, 14.9 id,  0.0 wa,  0.0 hi,  0.0 si,  0.0 st

Or use mpstat -P ALL

mpstat -P ALL
#01:23:14 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
#01:23:14 PM  all   33.33    0.00    8.61    0.04    0.00    3.00    0.00    0.00    0.00   55.02
#01:23:14 PM    0    2.46    0.00    1.32    0.06    0.00    6.27    0.00    0.00    0.00   89.88
#01:23:14 PM    1   61.65    0.00   15.29    0.02    0.00    0.00    0.00    0.00    0.00   23.03

Note: Use cat /proc/softirqs to check softirq type, please see Introduction to deferred interrupts (Softirq, Tasklets and Workqueues)

Note: If SRS run with softirq at CPU0, the total CPU will be larger than total of running on different CPUs.

If you got more CPUs, you can run softirq to multiple CPUs:

# grep virtio /proc/interrupts | grep -e in -e out
 29:   64580032          0          0          0   PCI-MSI-edge      virtio0-input.0
 30:          1         49          0          0   PCI-MSI-edge      virtio0-output.0
 31:   48663403          0   11845792          0   PCI-MSI-edge      virtio0-input.1
 32:          1          0          0         52   PCI-MSI-edge      virtio0-output.1

# cat /proc/irq/29/smp_affinity
1 # Bind softirq of virtio0 incoming to CPU0.
# cat /proc/irq/30/smp_affinity
2 # Bind softirq of virtio0 outgoing to CPU1.
# cat /proc/irq/31/smp_affinity
4 # Bind softirq of virtio1 incoming to CPU2.
# cat /proc/irq/32/smp_affinity
8 # Bind softirq of virtio1 outgoing to CPU3.

To disable softirq balance and force to run on CPU0, see Linux: scaling softirq among many CPU cores and SMP IRQ affinity by:

for irq in $(grep virtio /proc/interrupts | grep -e in -e out | cut -d: -f1); do 
    echo 1 > /proc/irq/$irq/smp_affinity
done

Note:Run echo 3 > /proc/irq/$irq/smp_affinity if bind to CPU0 and CPU1.

Then run SRS on other CPUs except CPU0:

taskset -a -p 0xfe $(cat objs/srs.pid)

You can improve about 20% performance by bind softirq to CPU0.

You can also setup in the startup script.

Process Priority

You can set SRS to run in higher priority:

renice -n -15 -p $(pidof srs)

Note: The value of nice is -20 to 19 and default is 0.

To check the priority, which is the NI field of top:

top -n1 -p $(pidof srs)
#  PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM     TIME+ COMMAND                
# 1505 root       5 -15  519920 421556   4376 S  66.7  5.3   4:41.12 srs

Performance Banchmark

The performance benchmark for SRS, compare with nginx-rtmp single process.

Provides detail benchmark steps.

The latest data, read performance.

Hardware

The client and server use lo net interface to test:

  • Hardware: VirtualBox on ThinkPad T430
  • OS: CentOS 6.0 x86_64 Linux 2.6.32-71.el6.x86_64
  • CPU: 3 Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz
  • Memory: 2007MB

OS

Login as root, set the fd limits:

  • Set limit: ulimit -HSn 10240
  • View the limit:
[root@dev6 ~]# ulimit -n
10240
  • Restart SRS:sudo /etc/init.d/srs restart

NGINX-RTMP

NGINX-RTMP version and build command.

  • NGINX: nginx-1.5.7.tar.gz
  • NGINX-RTMP: nginx-rtmp-module-1.0.4.tar.gz
  • Read nginx-rtmp
  • Build:
./configure --prefix=`pwd`/../_release \
--add-module=`pwd`/../nginx-rtmp-module-1.0.4 \
--with-http_ssl_module && make && make install
  • Config nginx:_release/conf/nginx.conf
user  root;
worker_processes  1;
events {
    worker_connections  10240;
}
rtmp{
    server{
        listen 19350;
        application live{
            live on;
        }
    }
}
  • The limit of fd:
[root@dev6 nginx-rtmp]# ulimit -n
10240
  • Start: ./_release/sbin/nginx
  • Check nginx started:
[root@dev6 nginx-rtmp]# netstat -anp|grep 19350
tcp        0      0 0.0.0.0:19350               0.0.0.0:*                   LISTEN      6486/nginx

SRS

SRS version and build.

  • SRS: SRS 0.9
  • Build: ./configure && make
  • Config SRS:conf/srs.conf
listen              1935;
max_connections     10240;
vhost __defaultVhost__ {
    gop_cache       on;
    forward         127.0.0.1:19350;
}
  • Check limit fds:
[root@dev6 trunk]# ulimit -n
10240
  • Start SRS: nohup ./objs/srs -c conf/srs.conf >/dev/null 2>&1 &
  • Check SRS started:
[root@dev6 trunk]# netstat -anp|grep "1935 "
tcp        0      0 0.0.0.0:1935                0.0.0.0:*                   LISTEN      6583/srs

Publish and Play

Use centos to publish RTMP:

  • Start FFMPEG:
for((;;)); do \
    ./objs/ffmpeg/bin/ffmpeg \
        -re -i doc/source.flv \
        -acodec copy -vcodec copy \
        -f flv rtmp://127.0.0.1:1935/live/livestream; \
    sleep 1; 
done
  • SRS RTMP stream URL: rtmp://192.168.2.101:1935/live/livestream
  • Nginx-RTMP stream URL: rtmp://192.168.2.101:19350/live/livestream

Client

The RTMP load test tool, read srs-bench

The sb_rtmp_load used to test RTMP load, support 800-3k concurrency for each process.

  • Build: ./configure && make
  • Start: ./objs/sb_rtmp_load -c 800 -r <rtmp_url>

Record Data

Record data before test:

  • Use top command:
srs_pid=$(pidof srs); \
nginx_pid=`ps aux|grep nginx|grep worker|awk '{print $2}'`; \
load_pids=`ps aux|grep objs|grep sb_rtmp_load|awk '{ORS=",";print $2}'`; \
top -p $load_pids$srs_pid,$nginx_pid
  • The connections:
srs_connections=`netstat -anp|grep srs|grep ESTABLISHED|wc -l`; \
nginx_connections=`netstat -anp|grep nginx|grep ESTABLISHED|wc -l`; \
echo "srs_connections: $srs_connections"; \
echo "nginx_connections: $nginx_connections";
  • The bandwidth in NBps:
[root@dev6 nginx-rtmp]# dstat -N lo 30
----total-cpu-usage---- -dsk/total- -net/lo- ---paging-- ---system--
usr sys idl wai hiq siq| read  writ| recv  send|  in   out | int   csw 
  0   0  96   0   0   3|   0     0 |1860B   58k|   0     0 |2996   465 
  0   1  96   0   0   3|   0     0 |1800B   56k|   0     0 |2989   463 
  0   0  97   0   0   2|   0     0 |1500B   46k|   0     0 |2979   461 
  • The table
ServerCPUMemConnENbpsANbpssbLat
SRS1.0%3MB3---0.8s
nginx-rtmp0.7%8MB2---0.8s

Memory(Mem): The memory usage in MB.

Clients(Conn): The connections/clients to server.

ExpectNbps(ENbps): The expect network bandwidth in Xbps.

ActualNbps(ANBps): The actual network bandwidth in Xbps.

srs-bench(srs-bench/sb): The mock benchmark client tool.

Latency(Lat): The latency of client.

Benchmark SRS

Let's start performance benchmark.

  • Start 500 clients
./objs/sb_rtmp_load -c 500 -r rtmp://127.0.0.1:1935/live/livestream >/dev/null &
  • The data:
ServerCPUMemConnENbpsANbpssbLat
SRS9.0%8MB503100Mbps112Mbps12.6%0.8s
  • The data for 1000 clients:
ServerCPUMemConnENbpsANbpssbLat
SRS23.6%13MB1003200Mbps239Mbps16.6%0.8s
  • The data for 1500 clients:
ServerCPUMemConnENbpsANbpssbLat
SRS38.6%20MB1503300Mbps360Mbps17%0.8s
  • The data for 2000 clients:
ServerCPUMemConnENbpsANbpssbLat
SRS65.2%34MB2003400Mbps480Mbps22%0.8s
  • The data for 2500 clients:
ServerCPUMemConnENbpsANbpssbLat
SRS72.9%38MB2503500Mbps613Mbps24%0.8s

Benchmark NginxRTMP

Let's start performance benchmark.

  • Start 500 clients:
./objs/sb_rtmp_load -c 500 -r rtmp://127.0.0.1:19350/live/livestream >/dev/null &
  • The data for 500 clients:
ServerCPUMemConnENbpsANbpssbLat
nginx-rtmp8.3%13MB502100Mbps120Mbps16.3%0.8s
  • The data for 1000 clients:
ServerCPUMemoryClientsExpectNbpsActualNbpssrs-benchLatency
nginx-rtmp27.3%19MB1002200Mbps240Mbps30%0.8s
  • The data for 1500 clients:
ServerCPUMemConnENbpsANbpssbLat
nginx-rtmp42.3%25MB1502300Mbps400Mbps31%0.8s
  • The data for 2000 clients:
ServerCPUMemConnENbpsANbpssbLat
nginx-rtmp48.9%31MB2002400Mbps520Mbps33%0.8s
  • The data for 2500 clients:
ServerCPUMemConnENbpsANbpssbLat
nginx-rtmp74.2%37MB2502500Mbps580Mbps35%0.8s

Performance Compare

ServerCPUMemConnENbpsANbpssbLat
nginx-rtmp8.3%13MB502100Mbps120Mbps16.3%0.8s
SRS9.0%8MB503100Mbps112Mbps12.6%0.8s
nginx-rtmp27.3%19MB1002200Mbps240Mbps30%0.8s
SRS23.6%13MB1003200Mbps239Mbps16.6%0.8s
nginx-rtmp42.3%25MB1502300Mbps400Mbps31%0.8s
SRS38.6%20MB1503300Mbps360Mbps17%0.8s
nginx-rtmp48.9%31MB2002400Mbps520Mbps33%0.8s
SRS65.2%34MB2003400Mbps480Mbps22%0.8s
nginx-rtmp74.2%37MB2502500Mbps580Mbps35%0.8s
SRS72.9%38MB2503500Mbps613Mbps24%0.8s

Performance Banchmark 4k

The performance is refined to support about 4k clients.

[winlin@dev6 srs]$ ./objs/srs -v
0.9.130
top - 19:52:35 up 1 day, 11:11,  8 users,  load average: 1.20, 1.05, 0.92
Tasks: 171 total,   4 running, 167 sleeping,   0 stopped,   0 zombie
Cpu0  : 26.0%us, 23.0%sy,  0.0%ni, 34.0%id,  0.3%wa,  0.0%hi, 16.7%si,  0.0%st
Cpu1  : 26.4%us, 20.4%sy,  0.0%ni, 34.1%id,  0.7%wa,  0.0%hi, 18.4%si,  0.0%st
Cpu2  : 22.5%us, 15.4%sy,  0.0%ni, 45.3%id,  1.0%wa,  0.0%hi, 15.8%si,  0.0%st
Mem:   2055440k total,  1972196k used,    83244k free,   136836k buffers
Swap:  2064376k total,     3184k used,  2061192k free,   926124k cached

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                                                                          
17034 root      20   0  415m 151m 2040 R 94.4  7.6  14:29.33 ./objs/srs -c console.conf                                                        
 1063 winlin    20   0  131m  68m 1336 S 17.9  3.4  54:05.77 ./objs/sb_rtmp_load -c 800 -r rtmp://127.0.0.1:1935/live/livestream               
 1011 winlin    20   0  132m  68m 1336 R 17.6  3.4  54:45.53 ./objs/sb_rtmp_load -c 800 -r rtmp://127.0.0.1:1935/live/livestream               
18736 winlin    20   0  113m  48m 1336 S 17.6  2.4   1:37.96 ./objs/sb_rtmp_load -c 800 -r rtmp://127.0.0.1:1935/live/livestream               
 1051 winlin    20   0  131m  68m 1336 S 16.9  3.4  53:25.04 ./objs/sb_rtmp_load -c 800 -r rtmp://127.0.0.1:1935/live/livestream               
18739 winlin    20   0  104m  39m 1336 R 15.6  2.0   1:25.71 ./objs/sb_rtmp_load -c 800 -r rtmp://127.0.0.1:1935/live/livestream   
[winlin@dev6 ~]$ dstat -N lo 30
----total-cpu-usage---- -dsk/total- ---net/lo-- ---paging-- ---system--
usr sys idl wai hiq siq| read  writ| recv  send|  in   out | int   csw 
  3   2  92   0   0   3|  11k   27k|   0     0 |   1B   26B|3085   443 
 32  17  33   0   0  17| 273B   60k|  69M   69M|   0     0 |4878  6652 
 34  18  32   0   0  16|   0    38k|  89M   89M|   0     0 |4591  6102 
 35  19  30   0   0  17| 137B   41k|  91M   91M|   0     0 |4682  6064 
 33  17  33   0   0  17|   0    31k|  55M   55M|   0     0 |4920  7785 
 33  18  31   0   0  17|2867B   34k|  90M   90M|   0     0 |4742  6530 
 32  18  33   0   0  17|   0    31k|  66M   66M|   0     0 |4922  7666 
 33  17  32   0   0  17| 137B   39k|  65M   65M|   0     0 |4841  7299 
 35  18  30   0   0  17|   0    28k| 100M  100M|   0     0 |4754  6752 
 32  17  33   0   0  18|   0    41k|  44M   44M|   0     0 |5130  8251 
 34  18  32   0   0  16|   0    30k| 104M  104M|   0     0 |4456  5718 

SRS 4k

Performance Banchmark 6k

SRS2.0.15, not SRS1.0, performance is refined to support 6k clients. That is 4Gbps for 522kbps bitrate, for a single SRS process. Read https://github.com/ossrs/srs/issues/194

Performance Banchmark 7.5k

SRS2.0.30 refined to support 7.5k clients, read https://github.com/ossrs/srs/issues/217

Winlin 2014.11