Tuning TCP/IP buffer sizes

WebSphere® Application Server uses the TCP/IP sockets communication mechanism extensively. For a TCP/IP socket connection, the send and receive buffer sizes define the receive window. The receive window specifies the amount of data that can be sent and not received before the send is interrupted. If too much data is sent, it overruns the buffer and interrupts the transfer. The mechanism that controls data transfer interruptions is referred to as flow control. If the receive window size for TCP/IP buffers is too small, the receive window buffer is frequently overrun, and the flow control mechanism stops the data transfer until the receive buffer is empty.

• Tune the TCP/IP buffer sizes.
  1. Change the TCP/IP configuration.
     1. Run the Change TCP/IP Attribute, CHGTCPA command.
     2. View and change the buffer sizes by pressing F4 on the Change TCP/IP Attributes window. The buffer sizes are displayed as the TCP receive and send buffer sizes. Type new values and save your changes.
  2. Recycle TCP/IP, then monitor CPU and paging rates to determine if they are within recommended system guidelines.
• Tune the TCP/IP buffer sizes.
  1. First, ensure that you have defined enough sockets to your system and that the default socket timeout of 180 seconds is not too high. To allow enough sockets, update the BPXPRMxx parmlib member:
     1. Set MAXSOCKETS for the AF_INET filesystem high enough.
     2. Set MAXFILEPROC high enough.

     Best practice: You should set MAXSOCKETS and MAXFILEPROC to at least 5000 for low-throughput, 10000 for medium-throughput, and 35000 for high-throughput WebSphere transaction environments. Setting high values for these parameters should not cause excessive use of resources unless the sockets or files are actually allocated.

     The following example shows settings for the MAXSOCKETS and MAXFILEPROC parameters:

     /* Open/MVS Parmlib Member                                           */
     /* CHANGE HISTORY:                                                   */
     /*   01/31/02 AEK Increased MAXSOCKETS on AF_UNIX from 10000 to 50000*/
     /*                per request from My Developer                      */
     /*   10/02/01 JAB Set up shared HFS                                  */
     
     /* KERNEL RESOURCES              DEFAULT             MIN MAX         */
     /* ========================      =================== === =========== */
        .
        .
        MAXFILEPROC(65535)            /* 64               3   65535       */
     
        .
        .
         NETWORK DOMAINNAME(AF_INET) DOMAINNUMBER(2) MAXSOCKETS(30000)
        .
     

  2. Next check the specification of the port in TCPIP profile dataset to ensure that NODELAYACKS is specified as follows:

     PORT 8082 TCP NODELAYACKS
     
     

     In your runs, changing this could improve throughput by as much as 50% (this is particularly useful when dealing with trivial workloads). This setting is important for good performance when running SSL.

  3. You should ensure that your DNS configuration is optimized so that lookups for frequently-used servers and clients are being cached.

     Caching is sometimes related to the name server's Time To Live (TTL) value. On the one hand, setting the TTL high will ensure good cache hits. However, setting it high also means that, if the Daemon goes down, it will take a while for everyone in the network to be aware of it.

     A good way to verify that your DNS configuration is optimized is to issue the oping and onslookup USS commands. Make sure they respond in a reasonable amount of time. Often a misconfigured DNS or DNS server name will cause delays of 10 seconds or more.

  4. Increase the size of the TCPIP send and receive buffers from the default of 16K to at least 64K. This is the size of the buffers including control information beyond what is present in the data that you are sending in your application. To do this specify the following:

     
     TCPCONFIG TCPSENDBFRSIZE 65535
               TCPRCVBUFRSIZE 65535
     
     

     Avoid trouble: It is unreasonable, in some cases, to specify 256 KB buffers.
  5. Increase the default listen backlog.
     The default listen backlog is used to buffer spikes in new connections which come with a protocol like HTTP. The default listen backlog is 10 requests. You should use the TCP transport channel listenBacklog custom property to increase this value to something larger. For example:

     listenBacklog=100

     Note: The value you use for listenBacklog can be limited by the specification of the SOMAXCONN statement in the TCP/IP profile. If you use a listenBacklog value greater than the SOMAXCONN value, the listenBacklog value is not be used; the value for SOMAXCONN is used.

     IMPORTANT: If listenBacklog is not set for channel types of HTTP, HTTP SSL, IIOP and IIOP SSL, the listenBacklog is set from deprecated environment values: protocol_http_backlog, protocol_https_backlog, protocol_iiop_backlog, and protocol_iiop_backlog_ssl. If the associated deprecated environment value is not specified, a default of 10 is used.

     For channel types that are not HTTP, HTTP SSL, IIOP and IIOP SSL, the default for listenBacklog is 511.

  6. Reduce the finwait2 time.

     In the most demanding benchmarks you may find that even defining 65K sockets and file descriptors does not give you enough 'free' sockets to run 100%. When a socket is closed abnormally (for example, no longer needed) it is not made available immediately. Instead it is placed into a state called finwait2 (this is what shows up in the netstat -s command). It waits there for a period of time before it is made available in the free pool. The default for this is 600 seconds.

     Avoid trouble: Unless you have trouble using up sockets, you should leave this set to the default value.

     If you are using z/OS® Version 1.2 or later you can control the amount of time the socket stays in finwait2 state by specifying the following in the configuration file:

     
     FINWAIT2TIME 60
     
     

Results

Repeat this process until you determine the ideal buffer size.

What to do next

The TCP/IP buffer sizes are changed. Repeat this process until you determine the ideal buffer size.