Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Transmission Control Protocol Express

IP.com Disclosure Number: IPCOM000104567D
Original Publication Date: 1993-Mar-01
Included in the Prior Art Database: 2005-Mar-19
Document File: 4 page(s) / 110K

Publishing Venue

IBM

Related People

Jung, CC: AUTHOR

Abstract

Enhanced Clustered FORTRAN (ECF) for a cluster of ES/9000*s or ES/3090*s (running AIX/370* or MVS/ESA*) requires a high speed, reliable, point-to-point Clustered Fortran link or HiPPI. Transmission Control Protocol/Internet Protocol (TCP/IP) is currently used to transfer data over the fast link between two ES/9000s, but the throughput (one user space to another user space) is very low (maximum of 5-7 MB/sec in AIX/370), especially when passing large blocks of data. The reasons for degraded throughput is due to under-utilization of the hardware as follows:

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Transmission Control Protocol Express

      Enhanced Clustered FORTRAN (ECF) for a cluster of ES/9000*s or
ES/3090*s (running AIX/370* or MVS/ESA*) requires a high speed,
reliable, point-to-point Clustered Fortran link or HiPPI.
Transmission Control Protocol/Internet Protocol (TCP/IP) is currently
used to transfer data over the fast link between two ES/9000s, but
the throughput (one user space to another user space) is very low
(maximum of 5-7 MB/sec in AIX/370), especially when passing large
blocks of data.  The reasons for degraded throughput is due to
under-utilization of the hardware as follows:

1.  Multiple memory-to-memory copying.  There are six times of
    copying involved to send a data from one ES/9000 user space to
    another (three times on sender and another three times on
    receiver) for ECF processing on AIX/370, and it becomes 12 for a
    round trip.  Three times of copying include user level copying,
    from user to the kernel buffer (mbufs) as required by TCP, and
    from the kernel buffer to the hardware (HiPPI or CFL) buffer.
    User level copying is necessary in ECF as it is based on RPC and
    External Data Representation library (XDR), and XDR requires a
    copying.  The number copying is higher in MVS due to additional
    memory-to-memory copying required by TCP/IP software.  See Fig.
    1.

2.  TCP/IP Segmentation.  Another problem with TCP/IP is that it uses
    unsigned and signed short integers (2 bytes) to set the maximum
    window size (64K) and the maximum transmission unit (MTU) size
    (32K), respectively.  This limits TCP/IP to segment data in a
    maximum of 32K byte chunks (theoretical maximum), and at most two
    packets can be transmitted before waiting for an acknowledgement.
    This causes significant under-utilization of CFL hardware which
    can support a packet size up to 256 M Bytes and the problem
    becomes magnified with large data transfers.

3.  Redundant protocol processing.  Because TCP protocol processing
    is performed independent from underlying connection hardware,
    some of the protocol processing becomes redundant when the
    underlying connection hardware and device driver handle the
    redundant functions already.  Point-to-point CFL and HiPPI
    connection eliminates the need to perform flow control and
    sequencing, and minimizes the need of error control for data
    corruption or packet loss.

      TCP Express is an auxiliary protocol which solves the ECF
TCP/IP throughput problems in a simple and cost effective manner.
TCP Express provides a simple way of exploiting advanced underlying
hardware features, such as large memory, fast and reliable link, for
fast transmission of large data between two remote user spaces.  TCP
Express complements TCP...