How the TCP Push Bit Was Changed for Windows NT 3.5

MORE INFORMATION

The TCP protocol has a bit in every header called the PUSH bit. RFC 1122 section 4.2.2.2 contains the following text on the PUSH bit.

   4.2.2.2  Use of Push: RFC-793 Section 2.8

      When an application issues a series of SEND calls without
      setting the PUSH flag, the TCP MAY aggregate the data
      internally without sending it. Similarly, when a series of
      segments is received without the PUSH bit, a TCP MAY queue
      the data internally without passing it to the receiving
      application.

      The PUSH bit is not a record marker and is independent of
      segment boundaries. The transmitter SHOULD collapse
      successive PUSH bits when it packetizes data, to send the
      largest possible segment.

      A TCP MAY implement PUSH flags on SEND calls. If PUSH flags
      are not implemented, then the sending TCP: (1) must not
      buffer data indefinitely, and (2) MUST set the PUSH bit in
      the last buffered segment (i.e., when there is no more
      queued data to be sent).

      The discussion in RFC-793 on pages 48, 50, and 74
      erroneously implies that a received PUSH flag must be passed
      to the application layer. Passing a received PUSH flag to
      the application layer is now OPTIONAL.

      An application program is logically required to set the PUSH
      flag in a SEND call whenever it needs to force delivery of
      the data to avoid a communication deadlock. However, a TCP
      SHOULD send a maximum-sized segment whenever possible, to
      improve performance (see Section 4.2.3.4).

      DISCUSSION:
         When the PUSH flag is not implemented on SEND calls,
         i.e., when the application/TCP interface uses a pure
         streaming model, responsibility for aggregating any
         tiny data fragments to form reasonable sized segments
         is partially borne by the application layer.

         Generally, an interactive application protocol must set
         the PUSH flag at least in the last SEND call in each
         command or response sequence. A bulk transfer protocol
         like FTP should set the PUSH flag on the last segment
         of a file or when necessary to prevent buffer deadlock.

         At the receiver, the PUSH bit forces buffered data to be
         delivered to the application (even if less than a full
         buffer has been received). Conversely, the lack of a
         PUSH bit can be used to avoid unnecessary wakeup calls
         to the application process; this can be an important
         performance optimization for large timesharing hosts.
         Passing the PUSH bit to the receiving application allows
         an analogous optimization within the application.
				

The receive-side use of the PUSH bit was changed in Windows NT version 3.5. In Windows NT version 3.1, TCP/IP behaved as if the PUSH bit was always set by the sender. Consequently, it completed application recv() calls as soon as any data was available. In Windows NT version 3.5, the receive logic was optimized to use the PUSH bit. This reduces the number of times an application has to wake up for incoming data. However, when NT is communicating with a TCP/IP implementation (or application) that does not set the PUSH bit at appropriate times, performance can suffer.

Performance degrades because, when data arrives without the PUSH bit set, TCP holds onto the application's recv() while waiting for the rest of the data. Windows NT version 3.5 TCP/IP completes a recv() call when:

• Data arrives with the PUSH bit set.
  
  -or-
• The user recv() buffer is full.
  
  -or-
• 0.5 seconds have elapsed since any data arrived.

When the third complete test above is required to complete a recv() call, performance degrades.

The following example shows how the improved handling of the TCP PUSH bit can improve performance:

   A client-server pair is going to transfer 4096 bytes between them on
   Ethernet. The client application executes a send() for 4096 bytes.
   Windows NT TCP/IP breaks the 4096 bytes into three Ethernet packets:

      1460 bytes, no PUSH
      1460 bytes, no PUSH
      1176 bytes, PUSH is set

   At the same time, the server executes recv(), for 8192 bytes. When the
   first packet arrives, TCP puts the data into the user buffer. The same
   thing happens for the second packet since PUSH is still not set.
   However, when the third packet arrives, TCP completes the recv() because
   PUSH was set.
				

In this scenario, holding on to the recv() boosts performance because of the overhead implicit in each recv() call; instead of making three calls to recv(), the application makes only one, returning all 4096 bytes. If the client did not set the PUSH flag on the third packet, TCP would not send recv() for another 0.5 seconds. Because some TCP/IP implementations (and applications) never set PUSH, they tend to perform poorly with Windows NT version 3.5 TCP/IP.