Linux Tuning For SIP Routers – Part 4 (Networking)

.

Introduction

This is Part  4 of “Linux Tuning For SIP Routers” topic. In Part 1 i have talked about Interrupts and IRQ Tuning. In Part 2 i have talked about File System Tuning, Journaling File System, and Swappiness . In Part 3 i have talked about OOM killer, Private Momory, and Shared Memory Tuning. In this part i will talk about Network tuning.

Network Adapter Settings

To check the setting of your network card, type: “ethtool X” where X is the interface name. The tool “ethtool” is used to read and write the settings of the network card. To update the settings:

  • Find the name of the network interface: # ifconfig
  • Get the current settings: # ethtool X
  • Do your change using ethtool (See the man page: # man ethtool). For example to change the speed to 1000, and the mode to  full duplex: # ethtool -s p4p2 speed 1000 duplex full autoneg off (where p4p2 is my interface name).

Note: Changes must be supported by the network adapter, otherwise it will give an error message “Cannot set new settings: Invalid argument”

To make the changes permanent for p4p2, set the environment variable  ETHTOOL_OPTS:

  • If you are using bash : # vi /etc/sysconfig/network-scripts/ifcfg-p4p1  (note “ifcfg-p4p1” will be different)
  • Add/update this line: ETHTOOL_OPTS=speed 1000 duplex full
  • Restart the network service: # systemctl restart network.service

“txqueuelen” Parameter

The kernel parameter “txqueuelen” is the size of the transmission queue related to the interface. The default value is 1000 frames. The kernel stores the departing frames in this queue (the frames are not loaded onto the NIC’s buffer yet). Tuning this value is important to avoid loosing frames due to the lack of space in the transmission queue. Use high value (long queue) for high speed interfaces (≈> 10 Gb).

To get the current value of txqueuelen: # ifconfig p4p2 where p4p2 is the interface name. The output will look like:

 p4p2: flags=4099<UP,BROADCAST,MULTICAST>  mtu 1500 ether 0c:54:a5:08:45:a7  txqueuelen 1000  (Ethernet) RX packets 0  bytes 0 (0.0 B) RX errors 0  dropped 0  overruns 0  frame 0 TX packets 0  bytes 0 (0.0 B) TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0 device interrupt 19

You can also execute “# ip link” to get the txqueuelen value (qlen 1000).

To temporarily change the value to 3000 (for ≈10 Gb card): # ifconfig  p4p2 txqueuelen 3000

To permanently change it, add/change the command “/sbin/ifconfig p4p2 txqueuelen 3000” in the file “/etc/rc.d/rc.local”. If this file is not exist, do the following if you have Fedora Red Hat:

  • # vi /etc/rc.d/rc.local
  • # chmod a+x /etc/rc.d/rc.local
  • At the top of the file, add the bash interpreter: #!/bin/sh
  • At the end of the file add exit 0
  • Between the previous two lines add the command: /sbin/ifconfig p4p2 txqueuelen 3000
  • Save the file and reboot (The command will be executed at system startup).

“netdev_max_backlog” Parameter

The kernel parameter “netdev_max_backlog” is the maximum size of the receive queue. The received frames will be stored in this queue after taking them from the ring buffer on the NIC. Use high value for high speed cards to prevent loosing packets. In real time application like SIP router, long queue must  be assigned with high speed CPU otherwise the data in the queue will be out of date (old).

To get the current value of netdev_max_backlog: #  /proc/sys/net/core/netdev_max_backlog. The default value is 1000 frames.

To temporarily change the size to 300000 frame:  # echo 300000 > /proc/sys/net/core/netdev_max_backlog

To permanently change the value:

  • Edit the file “/etc/sysctl.conf
  • Add the line: net.core.netdev_max_backlog=300000
  • Load the change which you did:  # sysctl -p

Note

  • Do a test before going to the deployment.
  • To change the ring buffer size on the NIC: # ethtool -g p4p2 where p4p2 is the interface name. This operation must be supported by the NIC.

Kernel Parameters For the Socket Buffer

Usually SIP uses TCP or UDP to carry the SIP signaling messages over the internet (<=> TCP/UDP sockets). The receive buffer (socket receive buffer) holds the received data until it is read by the application. The send buffer (socket transmit buffer) holds the data until it is read by the underling protocol in the network stack.

The default (initial) and maximum size (In  bytes) of the receive socket buffer:

# cat /proc/sys/net/core/rmem_default

# cat /proc/sys/net/core/rmem_max

You can manipulate this value in the application by calling the function “setsockopt” and specify the option name “SO_RCVBUF”  with the socket level set to SOL_SOCKET. When you specify a value for this option, the kernel will doubles this value.  So the doubled value should be less than the maximum (rmem_max).

The default (initial) and maximum size (In bytes) of the send socket buffer:

#  cat /proc/sys/net/core/wmem_default #  cat /proc/sys/net/core/wmem_max

You can manipulate this value in the application by calling the function “setsockopt” with the option name “SO_SNDBUF”. Choose the value which is suitable to he speed of the network card. For example, to set the maximum to 10 MB (10 * 1024 * 1024 Bytes)=10,485,760 for the receive buffer and send buffer:

Temporarily:

# echo 10485760 > /proc/sys/net/core/rmem_max

# echo 10485760 > /proc/sys/net/core/wmem_max

Permanently:

  • Edit the file “/etc/sysctl.conf
  • Add the lines:

net.core.rmem_max=10485760

net.core.wmem_max=10485760

  • Save the file and reload the changes:  # sysctl -p

To change the defaults:

Temporarily:

# echo 10485760 > /proc/sys/net/core/rmem_default

# echo 10485760 > /proc/sys/net/core/wmem_default

Permanently:

  • Edit the file “/etc/sysctl.conf
  • Add the lines:

net.core.rmem_default=10485760

net.core.wmem_default=10485760

  • Save the file and reload the changes:  # sysctl -p

Also You need to set the minimum size, initial size, and maximum size for the protocols. To read the current values (Assuming IPv4 and TCP):

# cat /proc/sys/net/ipv4/tcp_rmem

The output look like this: 4096    87380   6291456 (min, initial, max). To read write buffer settings: # cat /proc/sys/net/ipv4/tcp_wmem

You can do the same for UDP (udp_rmem and udp_wmem) To change the values: Temporarily:

# echo 10240 87380 10485760 > /proc/sys/net/ipv4/tcp_rmem

# echo 10240 87380 10485760 > /proc/sys/net/ipv4/tcp_rmem

Permanently:

  • Edit the file “/etc/sysctl.conf
  • Add the lines:

net.ipv4.tcp_rmem= 10240 87380 10485760

net.ipv4.tcp_wmem= 10240 87380 10485760

  • Save the file and reload the changes:  # sysctl -p

TCP Kernel Parameters

  • Disable TCP timestamp (RFC 1321): TCP timestamp feature allows round trip time measurement (<=>  Adding 8 bytes to TCP header). To avoid this overhead we disable this feature:  net.ipv4.tcp_timestamps = 0
  • Enable window scaling: net.ipv4.tcp_window_scaling = 1
  • Disable select acknowledgements (SACK): net.ipv4.tcp_sack = 1
  • Disable cache metrics so the initial conditions of the closed connections will not be saved to be used in near future connections:

net.ipv4.tcp_no_metrics_save = 1

‘1’ means disable caching.

  • Tune the value of “backlog” (maximum queue length of pending connections “Waiting Acknowledgment”):

tcp_max_syn_backlog= 300000

  • Set the value of somaxconn. This is the Max value of the backlog. The default value is 128. If the backlog is greater than somaxconn, it will truncated to it.

Temporarily# echo 300000 > /proc/sys/net/core/somaxconn

Permanently: Add the line: net.core.somaxconn= 300000 in the file /etc/sysctl.conf. Reload the change (# sysctl -p).

  • ” TIME_WAIT” TCP socket state is the state where the socket is closed but waiting to handle the packets which are still in the network. The parameter tcp_max_tw_buckets is the maximum number of sockets in “TIME_WAIT” state. After reaching this number the system will start destroying the socket in this state. To get the default value:

# cat /proc/sys/net/ipv4/tcp_max_tw_buckets.

Increase the value to 2000000 (2 million). Increasing this value leads to requirement of more memory.

Temporarily:

# echo 2000000 > /proc/sys/net/ipv4/tcp_max_tw_buckets

Permanently: Add the line: net.ipv4.tcp_max_tw_buckets= 2000000 in the file /etc/sysctl.conf. Reload the change (# sysctl -p).

  • Other TCP parameters that you can change as above:

net.ipv4.tcp_keepalive_time (how often the keepalive packets will be sent to keep the connection alive).

net.ipv4.tcp_keepalive_intvl (time to wait for a reply on each keepalive probe).

net.ipv4.tcp_retries2 (how many times to retry before killing an alive TCP connection).

net.ipv4.tcp_syn_retries (how many times to retransmit the initial SYN packet).


 Number of Open Files

When the SIP router is high loaded, this means it has a lot of open sockets (Each socket is a file). To increase the number of sockets we increase the number of open files (<=> number of file handles). To know the maximum number of file handles on the entire system:

# cat /proc/sys/fs/file-max. The output will be a number:  391884

To get the current usage of file handles: # cat /proc/sys/fs/file-nr. The output will look like: 10112   0       391884 First number (10112) is the total number of allocated file handles. Second number (0) is the currently used file handles (2.4 kkernel) or currently unused file handles (2.6 kernel). Third number (391884) is the maximum number of file handles.

To change the maximum number of open files temporarily:  # echo 65536 > /proc/sys/fs/file-max

To change the maximum number of open files permanently: Add the line “fs.file-max=65536”  in the file “/etc/sysctl.conf”.

SHELL Limits

The maximum number of open files “/proc/sys/fs/file-max” is for the whole system. This number is for all users so if you want to allow specific user (lets say opensips) to open X files where X < file-max, you can use “ulimit” to do that:

  • As root, edit the file: # vi /etc/security/limits.conf
  • Add/Change the following lines:

opensips              soft               nofile                   4096

opensips              hard               nofile                  40000

At the start time, the opensips can open till 4096 files (soft limit). If opensips gets an error message about running out of file handles, then opensips user can execute “ulimit -n New_Number” to increase the number of file handles to New_Number. The New_Number must be less than the hard limit (the second line) and the hard limit must be less than the file-max.

To avoid getting the error message (“out of file handles”), let opensips in its start scripts execute “ulimit -n New_Number”. Choosing the New_Number depends on your test to know how much your opensips needs. You can as root set permanently “soft limit=hard limit = New_Number”  in the file ” # vi /etc/security/limits.conf”.

Do not set the hard limit to a value equal to /proc/sys/fs/file-max, then the entire system might run out of file handles. Remember file-max is for the whole system.


 Note

  • Don’t use the numbers in this article. Do your tests on your system.
  • Use network performance testing tools like netperf, pktgen, and mpstat.
  • You always have to check how to optimize your hardware:
    • If the Network Interface Card support Flow Control (e.g e1000 network interface card).
    • Some NIC drivers support interrupt coalescing (multiple interrupts can be coalesced into one kernel interrupts). If your NIC support this, you can activate it as following: # ethtool -C p4p2 where p4p2 is the interface.

 More Information


Linux Tuning For SIP Routers – Part 3 (Memory)

.

Introduction

This is Part  3 of “Linux Tuning For SIP Routers” topic. In Part 1 i have talked about Interrupts and IRQ Tuning. In Part 2 i have talked about File System Tuning, Journaling File System, and Swappiness . In this part i will talk about OOM killer, Private Momory, and Shared Memory Tuning.


Out-Of-Memory (OOM) Killer

Linux kernel uses Out-Of-Memory (OOM) killing technique to recover the memory in “out-of-memory” situation. The OOM killer will select one or more processes to kill. You can make a specific process less likely to be killed by the OOM killer as following:

# echo -15 /proc/PID/oom_adj  where the PID is the process ID and oom_adj represents how likely is the process to be killed. oom_adj takes values between -16 to +15. Setting oom_adj to -17 exempts the process entirely from the OOM killer.

To know how likely is a process to be killed, view the content of the file “oom_score”:

# cat /proc/PID/oom_score


Private Memory (Per Process Memory) Settings

Check if there is any memory limitation for a user process. The command “ulimit -a” display the limitations on all resources not only the memory. Use the option -m to display the max memory size:

# ulimit -m

The output: unlimited

The command “ulimit” is built into the bash shell and used for read and write the limitations on the resources available to the shell and the processes started by it. The system must allow such control on the resources.

What remained is configuring the private memory of SIP router’s process. See the example below:

Example: Tuning the Private Memory of OpenSIPS Process

OpenSIPS private memory (file: mem/mem.h) is a memory dedicated to each process and used for process own business like building buffer for sending SIP message. The default size of private memory used by each OpenSIPS process is 1 MB. Common cases that need large private memory are loading user location contacts and doing NAT pinging. You can set the size of private memory (in megabytes) at startup by using the command parameter -M:

# ./opensips -M 10


Shared Memory Settings

The shared memory is memory shared between processes. The data that are stored  in the shared memory can be accessed by many processes. You need to care about shared memory if your SIP router is multi process and the processes/threads from different processes share some data. To see the shared memory settings in your system, type: # ipcs -lm

—— Shared Memory Limits ——–
max number of segments = 4096
max seg size (kbytes) = 32768
max total shared memory (kbytes) = 8388608
min seg size (bytes) = 1

Usually the SIP router has its own memory manager which allocates and frees the memory offered by the operating system between its processes. So tuning the memory involves tuning the memory at the operating system level and at the router level. Linux as an operating system has some parameters that you can configure:

SHMMAX Parameter

It is the maximum size (in bytes) of a single shared segment that a Linux process can allocate in its virtual address space. The value of this parameter depends on whether you are using 32 bit system or 64 bit system (size of the available address space).

What to do

  • Check your system (# uname -a): I am using Fedora Red Hat SMP kernel on 64 bit platform (X86-64) so the maximum size of a shared memory segment theoretically is 2^64bytes. This is correspond to all physical RAM that you have.
  • As a test case, I will set the value of the SHMMAX parameter to “1/4 of physical RAM”. I have 4 GB of RAM so I will set SHMMAX value to 1*1024*1024*1024=1,073,741,824 (bytes):
    • Temporarily:   echo 1073741824> /proc/sys/kernel/shmmax
    • Permanently: Edit the file “/etc/sysctl.conf” and set the entry: kernel.shmmax = 1073741824. This change can be reflected without restart by executing: “/sbin/sysctl -p“.

SHMALL Parameter

It is the total amount of shared memory pages that can be used system wide. Hence, SHMALL should always be at least SHMMAX / PAGE_SIZE. Execute “getconf PAGE_SIZE” to get the page size. In my case the page size is 4096 So the SHMALL value will be 1073741824/ 4096 = 262,144.

After configuring what the operating system allocate to your SIP router application, we need to configure the application itself. The following is an example:

Example: Tuning the Shared Memory of OpenSIPS SIP Router

OpenSIPS 1.X is pure multi process SIP server (version 2.x is multi threads). It stores the state of the calls (SIP transactions or SIP dialogs) in the shared memory accessed by all its processes (file: mem/shm_mem.h). Lets say the SHMEM_SIZE is the size of the shared memory. The recommended value of SHMEM_SIZE depends on the complexity of the routing script. When you run OpenSIPS, you can specify the entire shared memory (in MB) by using the command parameter -m. Ideally it is preferable to have SHMEM_SIZE equal to SHMMAX kernel parameter. This means OpenSIPS shared memory will fit in one shared memory segment. This leads to less job to memory manager and less memory errors. If  SHMEM_SIZE > SHMMAX, this means OpenSIPS memory manager will take more than one segment and manage access to them between all processes.
Here i will show you how to set OpenSIPS shared memory to 1 GB (SHMEM_SIZE = SHMMAX):

# ./opensips -m 1024 (1GB * 1024= 1024 MB)

  • Test the shared memory: You can get some statistics about the shared memory used and managed by OpenSIPS by sending FIFO commands to OpenSIPS. Have a look here on the SHMEM class. Use also system commands like “free” (MemShared or Shmem value) to get information about the memory.
  • The size of the recommended shared memory depends on the complexity of the routing script. So do the test many times for different values and choose the best for your system. Don’t forget to tune the value of  SHMMAX kernel parameter.

Note: Configure both the private memory (-M parameter) and the shared memory (-m parameter) at OpenSIPS startup. Test for different values and choose the one which suitable to your routing script.


Semaphores Settings

Semaphores are like counters that are used for controlling access by multiple processes to shared resources like shared memories. To see all semaphores settings: # ipcs -ls

—— Semaphore Limits ——–
max number of arrays = 128                      # SEMMNI Parameter
max semaphores per array = 250               # SEMMSL Parameter
max semaphores system wide = 32000      # SEMMNS Parameter
max ops per semop call = 32                     # SEMOPM Parameter
semaphore max value = 32767

or you can do this: # cat /proc/sys/kernel/sem

The output will be in this order (SEMMSL, SEMMNS, SEMOPM, SEMMNI):

250     32000   32      128

Tuning these values also depends on the complexity of the routing script. More complexity –> More loaded modules (shared libraries) –> more sharing –> more need for tuning the synchronization between processes.

Keep these values AS IS  if you don’t have problems with performance. To change these values temporarily:

# echo 250 32000 100 128 > /proc/sys/kernel/sem  

OR 

# sysctl -w kernel.sem=”250 32000 100 128″

To permanently change these values, Edit the file “/etc/sysctl.conf” and override the settings: kernel.sem=250 32000 100 128

Check The System Limits using “ulimit” System Property

Follow these steps to monitor the resources offered by the system to specific process:

  • Get the ID of the proess using # ps -A | grep …
  • Execute this: # cat /proc/$ProcessID/limits. Replace the ProcessID with the ID you get using the “ps” above.

The output will look like this:

Limit                     Soft Limit           Hard Limit           Units
Max cpu time              unlimited            unlimited            seconds
Max file size             unlimited            unlimited            bytes
Max data size             unlimited            unlimited            bytes
Max stack size            8388608              unlimited            bytes
Max core file size        0                    unlimited            bytes
Max resident set          unlimited            unlimited            bytes
Max processes             511830               511830               processes
Max open files            1024                 1024                 files
Max locked memory         65536                65536                bytes
Max address space         unlimited            unlimited            bytes
Max file locks            unlimited            unlimited            locks
Max pending signals       511830               511830               signals
Max msgqueue size         819200               819200               bytes
Max nice priority         0                    0
Max realtime priority     0                    0
Max realtime timeout      unlimited            unlimited            us


Next

After tuning the memory of the operating system, you need to care about the memory of  the SIP application itself. Here is an example (OpenSIPS Memory Debugging) .In the Next Part, I will talk about network tuning to optimize SIP Routers.