Network Working Group R. Braden
Request for Comment: 468 UCLA/CCN
NIC: 14742 March 8, 1973
FTP DATA COMPRESSION
Major design objectives of the proposed File Transfer Protocol (FTP)
are reliability and efficiency for transmission of large files.
Efficiency has two faces: efficiency of the host CPU's, and efficient
use of the Network bandwidth. Block mode is intended to minimize CPU
overhead for bandwidth efficiency, there is a mode called "HASP" in
RFC 454. The "HASP" mode of FTP is really transmission with data
compression, i.e., an encoding scheme to reduce the information
redundancy in the messages.
RFC 454 contains no explicit definition of the "HASP" or compressed
mode, but instead notes that a future RFC by yours truly will define
the mode. Students of FTP may find this scarcely credible, but you
are now reading the promised RFC. It turned out to be much farther
in the future than any of us expected. Mea Culpa.
In the early years of the Network, its major uses have been remote
terminal interactions and the small-to-medium-sized file transmission
typical of remote job entry. As facilities such as the Illiac IV and
the Data Machine become operational on the Network, and the Network
community begins to include users with heavy data transmission
requirements, large file transmission will become a major mode of
Network use. For example, one user of CCN expects to send 2 x 10**8
bits of data _each_ _day_ over the Network.
Local byte compression of the type proposed here is particular
effective for reducing the size of "printer" files such as those
transmitted under the Network RJE protocol. Experience with CCN's
RJS service has shown a typical compression of print files by a
factor of between two and three. Since FTP was intended to contain
the data transfer part of Network RJE protocol as a subset, it is
appropriate to include a print file compression mechanism in FTP.
These considerations led the FTP committee to include a compressed
mode within FTP.
The two main arguments for data compression are economics and
convenience (usability). Consider first economics, which is
essentially a trade-off between CPU time and transmission costs. Of
course, as long as Network use is a free commodity, the economics of
data compression are all bad. That happy state won't last forever.
What does data compression cost?
Let us consider only simple linear compression schemes, such as the
one proposed here. By linear, I mean that the CPU time to examine a
source record is proportional to number of bytes in the record. A
simple linear scheme could detect repeated single characters, for
example. One could imagine quadratic schemes, which detected
repeated substrings; but except for possible special circumstance
where the source stings have some structure known to the compression
algorithm, the CPU economics don't favor quadratic compression.
Assuming a reasonable figure for large-scale CPU costs in the
generation of CCN's 360/91, we concluded that an upper bound on CPU
costs for total compression and decompression would be 5 cents per
megabit; this is based on very loose coding of a simple linear
algorithm. This may be compared with the projected Network
transmission costs of over 30 cents per megabit (possibly a lot
Thus, the CPU time to conserve bandwidth costs significantly less
than the bandwidth saved. Both CPU costs and bandwidth costs are
trending downward, but it seems exceedingly unlikely that the ratio
of CPU cost to bandwidth cost for linear compression will reverse in
the next few years. On the other hand, this calculation clearly
discourages one from using quadratic compression.
CCN's batch remote job entry protocol NETRJS (see RFC #189, July 15,
1971) was designed to include two data transfer modes, truncated and
compressed. The NETRJS truncated mode is essentially identical to
current FTP block mode record structure (except for minor bit format
differences). The compressed mode of NETRJS uses an adaptation of
the particular compression scheme which is incorporated in the
"Multileaving protocol" of the binary synchronous rje support in
IBM's HASP system.
Although it isn't really necessary for the purpose of defining a
compression scheme in FTP, I have included an appendix summarizing
very briefly the nature of HASP and its rje package. That appendix
may be considered cultural enrichment for those in the Network
Community who have been denied the privilege of being an IBM
customer. After all, I know a lot of HASP experts who never heard of
TENEX! More seriously, because HASP is widely used on IBM machines,
the HASP compression scheme is also widely used. In designing
NETRJS, we chose the HASP scheme of compression because of its
ubiquity and plausibility.
However, certain details of the HASP bit formats are inappropriate or
sub-optimal for FTP. Therefore, our proposal for compressed mode of
FTP is only an adaptation of the HASP compression scheme.
It should be clear from Appendix A that the compression scheme of
HASP, even if used literally, is a very minor and incidental part of
that system. Although we ought to properly credit the intellectual
origin of FTP's compressed mode, it seems a little strange to call
the compressed mode in FTP the "HASP mode". I trust this will be
rectified by the forthcoming FTP meeting.
II. PROPOSED FTP COMPRESSED DATA MODE
Byte size is B bits. Figures above boxes are field lengths in bits.
n bytes of data
1 B-1 / B B \
+---+------+ +--------+ +--------+
Byte String: | 0 | n | | d |. . .| d |
| | | | 1 | | n |
+---+------+ +--------+ +--------+
String of n data bytes d(1),...,d(n)
Count n must be positive
2 B-2 B
Replicated Byte: | 1 0| n | | d |
String consisting of n replications of the data byte d
Filler String: | 1 1| n |
String of n filler bytes. The filler byte is a "space"
character for ASCII or EBCDIC type, or a binary zero
byte for Image or Local Byte Type.
Control Escape Sequence: | 0......0 | | C | (see below)
The control byte "C" which is the second byte of a control escape
sequence is to have the same coding as the descriptor byte in Block
Mode. This includes end-of-file and end-of-record indications. I
will not specify this further because there is some question at
present about the exact coding of the Block Mode descriptor byte.
Following the example of APL*, we have let the meaning of the filler
(blank or 0) be determined by the type: character (ASCII|EBCDIC) vs.
binary (Image|Local Byte). If byte size is equal to the word size of
the transmitting host, the compressed mode allows one to send sparse
notices with reasonable efficiency.
* Compare 1 (take) 0 1\`A' with 1 (take) 0 1\2
APPENDIX A: HASP MULTILEAVING
HASP (Houston Automatic Spooling Program) is a subsystem which
essentially runs within OS/360 as a job but takes over the batch
processing management functions from the operating system. That is,
HASP handles spooling of card input and printer and punch output,
queueing and scheduling of job execution, and the operator control
interface. It is a tightly-written and efficient system for running
a large and varied job load through a large-scale machine. The name
results from the historical fact that HASP was originally by a local
IBM group for one particular customer, NASA Houston.
HASP has always been an anomaly in the IBM scheme of things. The
system was written around 1965 by two programmers; the HASP group has
probably averaged three programmers during most of its life. The
leader of the group has been "Mr. HASP", Tom Simpson. The HASP
system spread rapidly through (more or less) underground channels to
many of the medium and large scale 360's. At least once, only
intense customer pressure prevented IBM from killing the HASP effort.
HASP generated an astonishing emotional mystique among its users.
The HASP sessions at SHARE Meetings were reminiscent of revival
meetings. For years every SHARE Meeting has included HASP song
sessions around the piano during the nightly open bar. HASP forms a
fascinating chapter in the history of IBM's large machine business.
The core concepts in HASP are pseudo-devices, and the general
technique of intercepting supervisor calls to augment operating
system functions without changing the operating system itself. A
generation of OS/360 system programmers learned these techniques from
HASP. (These important techniques are hardly ever described in the
literature, and "practical" programmers don't read the literature
When HASP starts up (in supervisor state), it overlays an instruction
in the I/O Supervisor with a branch to its own code. A user program
is written as if it were doing real I/O to card readers and printers.
HASP intercepts and interprets these I/O operations to handle job I/O
in a manner transparent to OS/360. It similarly intercepts and
interprets operator console I/O.
HASP includes batch remote job entry using binary synchronous
communication. The HASP communication protocol and message formats
use a scheme developed by Simpson's group called "Multileaving
Protocol". The HASP rje system, by far the best rje package IBM has
produced, finally replaced two competitive IBM packages and has
effectively become the IBM standard for rje. CCN's RJS system not
only adopted the Multileaving Protocol but essentially copied its
binary synchronous communication line handler directly form HASP.
The Multileaving Protocol is described in the HASP manual(1) as the
"fully synchronized, pseudo-simultaneous, bidirectional transmission
of a variable number of data streams between two or more computers
using binary synchronous communications facilities". It allows a
remote batch terminal to operate a variable number of card readers
and printers simultaneously at different speeds over one
communication line. It is not surprising that HASP Multileaving
contains in miniature many of the features of IMP-IMP Protocol and a
little host-host protocol. Specifically, Multileaving includes the
following general features:
(1) "Conversational" transmission line protocol using transparency
(DLE STX, etc.).
(2) "Strong" error control and retransmission using a 16-bit CRC
and a modulo-16 block sequence number.
(3) Flow control for multiple streams in both directions. This
includes the interchanging of matching control records
("RFC's") to open a stream, and a set of flow control bits in
each block. Each flow control bit is logically equivalent to
an ALLOcate command for one "message" (buffer) for a
(4) Optional Special Control Information for remote devices. This
includes printer carriage control, switching card reader
(5) Multiplexing ("multileaving") multiple streams into a single
block for transmission.
(6) Marking end of file and ends of records within each stream.
(7) Compressing transmitted text by encoding repeated blanks and
repeated single characters.
Finally, we have reached the (only) aspect of HASP relevant to FTP:
its compression scheme. HASP uses the following encoding:
End of Record: | 0 ... 0 |
2 6 8 8
+---+---------+ +-------+ +--------+
Data String: |1 1| N | | d | ... | d |
| | | | 1 | | N |
+---+---------+ +-------+ +--------+
N Duplicate Blanks |100| N |
3 5 8
N Replicated Characters D |101| N | | D |
HASP is concerned only with 8-bit bytes. However, there is a
provision (which was never implemented) in the Multileaving Protocol
to set the unit of the counts N as 1 byte, 2 bytes, or 4 bytes.
(1) HASP II System Manual, IBM Corporation (February 26, 1971)
[ This RFC was put into machine readable form for entry ]
[ into the online RFC archives by Via Genie 4/00]