Network Working Group W. Polites
Request for Comments: 1986 W. Wollman
Category: Experimental D. Woo
The MITRE Corporation
R. Langan
U.S. ARMY CECOM
August 1996
Experiments with a Simple File Transfer Protocol for Radio Links
using Enhanced Trivial File Transfer Protocol (ETFTP)
Status of this Memo
This memo defines an Experimental Protocol for the Internet
community. This memo does not specify an Internet standard of any
kind. Discussion and suggestions for improvement are requested.
Distribution of this memo is unlimited.
This document is a description of the Enhanced Trivial File Transfer
Protocol (ETFTP). This protocol is an experimental implementation of
the NETwork BLock Transfer Protocol (NETBLT), RFC 998 [1], as a file
transfer application program. It uses the User Datagram Protocol
(UDP), RFC 768 [2], as its transport layer. The two protocols are
layered to create the ETFTP client server application. The ETFTP
program is named after Trivial File Transfer Protocol (TFTP), RFC
1350 [3], because the source code from TFTP is used as the building
blocks for the ETFTP program. This implementation also builds on but
differs from the work done by the National Imagery Transmission
Format Standard [4].
This document is published for discussion and comment on improving
the throughput performance of data transfer utilities over Internet
Protocol (IP) compliant, half duplex, radio networks.
There are many file transfer programs available for computer
networks. Many of these programs are designed for operations through
high-speed, low bit error rate (BER) cabled networks. In tactical
radio networks, traditional file transfer protocols, such as File
Transfer Protocol (FTP) and TFTP, do not always perform well. This is
primarily because tactical half duplex radio networks typically
provide slow-speed, long delay, and high BER communication links.
ETFTP is designed to allow a user to control transmission parameters
to optimize file transfer rates through half-duplex radio links.
Polites, Wollman & Woo Experimental [Page 1]
RFC 1986 ETFTP August 1996
The tactical radio network used to test this application was
developed by the Survivable Adaptive Systems (SAS) Advanced
Technology Demonstration (ATD). Part of the SAS ATD program was to
address the problems associated with extending IP networks across
tactical radios. Several tactical radios, such as, SINgle Channel
Ground and Airborne Radio Systems (SINCGARS), Enhanced Position
Location Reporting Systems (EPLRS), Motorola LST-5C, and High
Frequency (HF) radios have been interfaced to the system. This
document will discuss results obtained from using ETFTP across a
point-to-point LST-5C tactical SATellite COMmunications (SATCOM)
link. The network includes a 25 Mhz 486 Personal Computer (PC) called
the Army Lightweight Computer Unit (LCU), Cisco 2500 routers,
Gracilis PackeTen Network switches, Motorola Sunburst Cryptographic
processors, a prototype forward error correction (FEC) device, and
Motorola LST-5C tactical Ultra High Frequency (UHF) satellite
communications (SATC! OM) radio. Table 1, "Network Trans fer Rates,"
describes the equipment network connections and the bandwidth of the
physical media interconnecting the devices.
Table 1: Network Transfer Rates
+-------------------------------+-------------------------------+
| Equipment | Rate (bits per second) |
+-------------------------------+-------------------------------+
| Host Computer (486 PC) | 10,000,000 Ethernet |
+-------------------------------+-------------------------------+
| Cisco Router | 10,000,000 Ethernet to |
| | 19,200 Serial Line Internet |
| | Protocol (SLIP) |
+-------------------------------+-------------------------------+
| Gracilis PackeTen | 19,200 SLIP to |
| | 16,000 Amateur Radio (AX.25) |
+-------------------------------+-------------------------------+
| FEC | half rate or quarter rate |
+-------------------------------+-------------------------------+
| Sunburst Crypto | 16,000 |
+-------------------------------+-------------------------------+
| LST-5C Radio | 16,000 |
+-------------------------------+-------------------------------+
During 1993, the MITRE team collected data for network configurations
that were stationary and on-the-move. This network configuration did
not include any Forward Error Correction (FEC) at the link layer.
Several commercially available implementations of FTP were used to
transfer files through a 16 kbps satellite link. FTP relies upon the
Transmission Control Protocol (TCP) for reliable communications. For
a variety of file sizes, throughput measurements ranged between 80
and 400 bps. At times, TCP connections could be opened, however, data
Polites, Wollman & Woo Experimental [Page 2]
RFC 1986 ETFTP August 1996
transfers would be unsuccessful. This was most likely due to the
smaller TCP connection synchronization packets, as compared to the
TCP data packets. Because of the high bit error rate of the link,
the smaller packets were much more likely to be received without
error. In most cases, satellite channel utilization was less than 20
percent. Very often a file transfer would fail because FTP
implementations would curtail the transfer due t! o the poor
conditions of the commu nication link.
The current focus is to increase the throughput and channel
utilization over a point to point, half duplex link. Follow on
experiments will evaluate ETFTP's ability to work with multiple hosts
in a multicast scenario. Evaluation of the data collected helped to
determine that several factors limited data throughput. A brief
description of those limiting factors, as well as, solutions that can
reduce these networking limitations is provided below.
Link Quality
The channel quality of a typical narrow-band UHF satellite link does
not sufficiently support data communications without the addition of
a forward error correction (FEC) capability. From the data
collected, it was determined that the UHF satellite link supports, on
average, a 10e-3 bit error rate.
Solution: A narrow-band UHF satellite radio FEC prototype was
developed that improves data reliability, without excessively
increasing synchronization requirements. The prototype FEC increased
synchronization requirements by less than 50 milliseconds (ms). The
FEC implementation will improve an average 10e-3 BER channel to an
average 10e-5 BER channel.
Delays
Including satellite propagation delays, the tactical satellite radios
require approximately 1.25 seconds for radio synchronization prior to
transmitting any data across the communication channel. Therefore,
limiting the number of channel accesses required will permit data
throughput to increase. This can be achieved by minimizing the number
of acknowledgments required during the file transfer. FTP generates
many acknowledgments which decreases throughput by increasing the
number of satellite channel accesses required.
To clarify, when a FTP connection request is generated, it is sent
via Ethernet to the router and then forwarded to the radio network
controller (RNC). The elapsed time is less than 30 ms. The RNC keys
the crypto unit and 950 ms later modem/crypto synchronization occurs.
After synchronization is achieved, the FTP connection request is
Polites, Wollman & Woo Experimental [Page 3]
RFC 1986 ETFTP August 1996
transmitted. The transmitting terminal then drops the channel and the
modem/crypto synchronization is lost. Assuming that the request was
received successfully, the receiving host processes the request and
sends an acknowledgment. Again the modem/crypto have to synchronize
prior to transmitting the acknowledgment. Propagation delays over a
UHF satellite also adds roughly 500 ms to the total round trip delay.
Solution: When compared to FTP, NETBLT significantly reduces the
number of acknowledgments required to complete a file transfer.
Therefore, leveraging the features available within an implementation
of NETBLT will significantly improve throughput across the narrow-
band UHF satellite communication link.
To reduce the number of channel accesses required, a number of AX.25
parameters were modified. These included the value of p for use
within the p-persistence link layer protocol, the slot time, the
transmit tail time, and the transmit delay time. The p-persistence
is a random number threshold between 0 and 255. The slot time is the
time to wait prior to attempting to access the channel. The transmit
tail increases the amount of time the radio carrier is held on, prior
to dropping the channel. Transmit delay is normally equal to the
value of the radio synchronization time. By adjusting these
parameters to adapt to the tactical satellite environment, improved
communication performance can be achieved.
First, in ETFTP, several packets within a buffer are transmitted
within one burst. If the buffer is partitioned into ten packets, each
of 1024 bytes, then 10,240 bytes of data is transmitted with each
channel access. It is possible to configure ETFTP's burstsize to
equal the number of packets per buffer. Second, the transmit tail
time was increased to hold the key down on the transmitter long
enough to insure all of the packets within the buffer are sent in a
single channel access. These two features, together, allow the system
to transmit an entire file (example, 100,000 bytes) with only a
single channel access by adjusting buffer size. Thirdly, the ETFTP
protocol only acknowledges each buffer, not each packet. Thus, a
single acknowledgment is sent from the receiving terminal containing
a request for the missing packets within each buffer, reducing the
number of acknowledgment packets sent. Which in turn, reduced the
number of times the channel has to be turned around.
To reduce channel access time, p-persistence was set to the maximum
value and slot time to a minimum value. These settings support
operations for a point-to-point communication link between two users.
This value of p would not be used if more users were sharing the
satellite channel.
Polites, Wollman & Woo Experimental [Page 4]
RFC 1986 ETFTP August 1996
Backoffs
TCP's slow start and backoff algorithms implemented in most TCP
packages assume that packet loss is due to network congestion. When
operating across a tactical half duplex communication channel
dedicated to two users, packet loss is primarily due to poor channel
quality, not network congestion. A linear backoff at the transport
layer is recommended. In a tactical radio network there are numerous
cases where a single host is connected to multiple radios. In a
tactical radio network, layer two will handle channel access.
Channel access will be adjusted through parameters like p-persistence
and slot time. The aggregate effect of the exponential backoff from
the transport layer added to the random backoff of the data link
layer, will in most cases, cause the radio network to miss many
network access opportunities. A linear backoff will reduce the number
missed data link network access opportunities
Solution: Tunable parameters and timers have been modified to
resemble those suggested by NETBLT.
Packet Size
In a tactical environment, channel conditions change rapidly.
Continuously transmitting large packets under 10e-3 BER conditions
reduces effective throughput.
Solution: Packet sizes are dynamically adjusted based upon the
success of the buffer transfers. If 99 percent of all packets within
a buffer are received successfully, packet size can be increased to a
negotiated value. If 50 percent or more of all packets within a
buffer are not successfully delivered, the packet size can be
decreased to a negotiated value.
Throughout this document the term packet is used to describe a
datagram that includes all network overhead. A block is used to
describe information, without any network encapsulation.
The original source files for TFTP, as downloaded from ftp.uu.net,
were modified to implement the ETFTP/NETBLT protocol. These same
files are listed in "UNIX Network Programming" [5].
ETFTP was implemented for operations under the Santa Cruz Operations
(SCO) UNIX. In the service file, "/etc/services", an addition was
made to support "etftp" at a temporary well known port of "1818"
using "UDP" protocol. The file, "/etc/inetd.conf", was modified so
the "inetd" program could autostart the "etftpd" server when a
Polites, Wollman & Woo Experimental [Page 5]
RFC 1986 ETFTP August 1996
connection request came in on the well known port.
As stated earlier, the transport layer for ETFTP is UDP, which will
not be discussed further here. This client server application layer
protocol is NETBLT, with four notable differences.
The first change is that this NETBLT protocol is implemented on top
of the UDP layer. This allowed the NETBLT concepts to be tested
without modifying the operating system's transport or network layers.
Table 2, "Four Layer Protocol Model," shows the protocol stack for
FTP, TFTP and ETFTP.
Table 2: Four Layer Protocol Model
+---------------------------------------------------------------+
| PROTOCOL STACK |
+---------------+---------------+---------------+---------------+
|APPLICATION |FTP |TFTP |ETFTP/NETBLT |
+---------------+---------------+---------------+---------------+
|TRANSPORT |TCP |UDP |UDP |
+---------------+---------------+---------------+---------------+
|NETWORK |IP |
+---------------+---------------+---------------+---------------+
|LINK |Ethernet, SLIP, AX.25 |
+---------------+---------------+---------------+---------------+
The second change is a carryover from TFTP, which allows files to be
transferred in netascii or binary modes. A new T bit flag is assigned
to the reserved field of the OPEN message type.
The third change is to re-negotiate the DATA packet size. This change
affects the OPEN, NULL-ACK, and CONTROL_OK message types. A new R
bit is assigned to the reserved field of the OPEN message type.
The fourth change is the addition of two new fields to the OPEN
message type. The one field is a two byte integer for radio delay in
seconds, and the next field is two bytes of padding.
The ETFTP data encapsulation is shown in Table 3, "ETFTP Data
Encapsulation,". The Ethernet, SLIP, and AX.25 headers are mutually
exclusive. They are stripped off and added by the appropriate
hardware layer.
Polites, Wollman & Woo Experimental [Page 6]
RFC 1986 ETFTP August 1996
Table 3: ETFTP Data Encapsulation
+------------+------------+------------+------------+-----------+
|Ethernet(14)| | |ETFTP/ | |
|SLIP(2) |IP(20) |UDP(8) |NETBLT(24) |DATA(1448) |
|AX.25(20) | | | | |
+------------+------------+------------+------------+-----------+
Here are the ETFTP/NETBLT message types and formats.
MESSAGES VALUES
OPEN 0 Client request to open a new connection
RESPONSE 1 Server positive acknowledgment for OPEN
KEEPALIVE 2 Reset the timer
QUIT 3 Sender normal Close request
QUITACK 4 Receiver acknowledgment of QUIT
ABORT 5 Abnormal close
DATA 6 Sender packet containing data
LDATA 7 Sender last data block of a buffer
NULL-ACK 8 Sender confirmation of CONTROL_OK changes
CONTROL 9 Receiver request to
GO 0 Start transmit of next buffer
OK 1 Acknowledge complete buffer
RESEND 2 Retransmit request
REFUSED 10 Server negative acknowledgment of OPEN
DONE 11 Receiver acknowledgment of QUIT.
Packets are "longword-aligned", at four byte word boundaries.
Variable length strings are NULL terminated, and padded to the four
byte boundary. Fields are listed in network byte order. All the
message types share a common 12 byte header. The common fields are:
Checksum IP compliant checksum
Version Current version ID
Type NETBLT message type
Length Total byte length of packet
Local Port My port ID
Foreign Port Remote port ID
Padding Pad as necessary to 4 byte boundary
The OPEN and RESPONSE messages are similar and shown in Table 4,
"OPEN and RESPONSE Message Types,". The client string field is used
to carry the filename to be transferred.
Polites, Wollman & Woo Experimental [Page 7]
RFC 1986 ETFTP August 1996
Table 4: OPEN and RESPONSE Message Types
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Checksum |Version |Type |
+---------------+---------------+---------------+---------------+
|Length |Local Port |
+---------------+---------------+---------------+---------------+
|Foreign Port |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
|Connection ID |
+---------------+---------------+---------------+---------------+
|Buffer size |
+---------------+---------------+---------------+---------------+
|Transfer size |
+---------------+---------------+---------------+---------------+
|DATA Packet size |Burstsize |
+---------------+---------------+---------------+---------------+
|Burstrate |Death Timer Value |
+---------------+---------------+---------------+---------------+
|Reserved(MBZ) |R|T|C|M|Maximum # Outstanding Buffers |
+---------------+---------------+---------------+---------------+
|*Radio Delay |*Padding |
+---------------+---------------+---------------+---------------+
|Client String . . . |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
Connection ID The unique connection number
Buffer size Bytes per buffer
Transfer size The length of the file in bytes
DATA Packet size Bytes per ETFTP block
Burstsize Concatenated packets per burst
Burstrate Milliseconds per burst
Death Timer Seconds before closing idle links
Reserved M bit is mode: 0=read/put, 1=write/get
C bit is checksum: 0=header, 1=all
*T bit is transfer: 0=netascii, 1=binary
*R bit is re-negotiate: 0=off, 1=on
Max # Out Buffs Maximum allowed un-acknowledged buffers
Radio Delay *Seconds of delay from send to receive
Padding *Unused
Client String Filename.
The KEEPALIVE, QUITACK, and DONE messages are identical to the common
header, except for the message type values. See Table 5, "KEEPALIVE,
QUITACK, and DONE Message Types,".
Polites, Wollman & Woo Experimental [Page 8]
RFC 1986 ETFTP August 1996
Table 5: KEEPALIVE, QUITACK, and DONE Message Types
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Checksum |Version |Type |
+---------------+---------------+---------------+---------------+
|Length |Local Port |
+---------------+---------------+---------------+---------------+
|Foreign Port |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
The QUIT, ABORT, and REFUSED messages allow a string field to carry
the reason for the message. See Table 6, "QUIT, ABORT, and REFUSED
Message Types,".
Table 6: QUIT, ABORT, and REFUSED Message Types
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Checksum |Version |Type |
+---------------+---------------+---------------+---------------+
|Length |Local Port |
+---------------+---------------+---------------+---------------+
|Foreign Port |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
|Reason for QUIT/ABORT/REFUSED . . . |
+---------------+---------------+---------------+---------------+
|. . . |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
The DATA and LDATA messages make up the bulk of the messages
transferred. The last packet of each buffer is flagged as an LDATA
message. Each and every packet of the last buffer has the reserved L
bit set. The highest consecutive sequence number is used for the
acknowledgment of CONTROL messages. It should contain the ID number
of the current CONTROL message being processed. Table 7, "DATA and
LDATA Message Types,", shows the DATA and LDATA formats.
Polites, Wollman & Woo Experimental [Page 9]
RFC 1986 ETFTP August 1996
Table 7: DATA and LDATA Message Types
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Checksum |Version |Type |
+---------------+---------------+---------------+---------------+
|Length |Local Port |
+---------------+---------------+---------------+---------------+
|Foreign Port |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
|Buffer Number |
+---------------+---------------+---------------+---------------+
|High Consecutive Seq Num Rcvd |Packet Number |
+---------------+---------------+---------------+---------------+
|Data Area Checksum Value |Reserved (MBZ) |L|
+---------------+---------------+---------------+---------------+
Buffer Number The first buffer number starts at 0
Hi Con Seq Num The acknowledgment for CONTROL messages
Packet Number The first packet number starts at 0
Data Checksum Checksum for data area only
Reserved L: the last buffer bit: 0=false, 1=true
The NULL-ACK message type is sent as a response to a CONTROL_OK
message that modifies the current packet size, burstsize, or
burstrate. In acknowledging the CONTROL_OK message, the sender is
confirming the change request to the new packet size, burstsize, or
burstrate. If no modifications are requested, a NULL-ACK message is
unnecessary. See Table 8, "NULL-ACK Message Type," for further
details.
Table 8: NULL-ACK Message Type
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Checksum |Version |Type |
+---------------+---------------+---------------+---------------+
|Length |Local Port |
+---------------+---------------+---------------+---------------+
|Foreign Port |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
|High Consecutive Seq Num Rcvd |New Burstsize |
+---------------+---------------+---------------+---------------+
|New Burstrate |*New DATA Packet size |
+---------------+---------------+---------------+---------------+
Polites, Wollman & Woo Experimental [Page 10]
RFC 1986 ETFTP August 1996
The CONTROL messages have three subtypes: GO, OK, and RESEND as shown
in Tables 9-12. The CONTROL message common header may be followed by
any number of longword aligned subtype messages.
Table 9: CONTROL Message Common Header
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Checksum |Version |Type |
+---------------+---------------+---------------+---------------+
|Length |Local Port |
+---------------+---------------+---------------+---------------+
|Foreign Port |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
Table 10: CONTROL_GO Message Subtype
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Subtype |Padding |Sequence Number |
+---------------+---------------+---------------+---------------+
|Buffer Number |
+---------------+---------------+---------------+---------------+
Table 11: CONTROL_OK Message Subtype
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Subtype |Padding |Sequence Number |
+---------------+---------------+---------------+---------------+
|Buffer Number |
+---------------+---------------+---------------+---------------+
|New Offered Burstsize |New Offered Burstrate |
+---------------+---------------+---------------+---------------+
|Current Control Timer Value |*New DATA Packet size |
+---------------+---------------+---------------+---------------+
Polites, Wollman & Woo Experimental [Page 11]
RFC 1986 ETFTP August 1996
Table 12: CONTROL_RESEND Message Subtype
1 2 3
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2
+---------------+---------------+---------------+---------------+
|Subtype |Padding |Sequence Number |
+---------------+---------------+---------------+---------------+
|Buffer Number |
+---------------+---------------+---------------+---------------+
|Number of Missing Packets |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
|Packet Number (2 bytes) |. . . |
+---------------+---------------+---------------+---------------+
|. . . |Longword Alignment Padding |
+---------------+---------------+---------------+---------------+
Being built from TFTP source code, ETFTP shares a significant portion
of TFTP's design. Like TFTP, ETFTP does NOT support user password
validation. The program does not support changing directories (i.e.
cd), neither can it list directories, (i.e. ls). All filenames must
be given in full paths, as relative paths are not supported. The
internal finite state machine was modified to support NETBLT message
types.
The NETBLT protocol is implemented as closely as possible to what is
described in RFC 998, with a few exceptions. The client string field
in the OPEN message type is used to carry the filename of the file to
be transferred. Netascii or binary transfers are both supported. If
enabled, new packet sizes, burstsizes, and burstrates are re-
negotiated downwards when half or more of the blocks in a buffer
require retransmission. If 99% of the packets in a buffer is
successfully transferred without any retransmissions, packet size is
re-negotiated upwards.
The interactive commands supported by the client process are similar
to TFTP. Here is the ETFTP command set. Optional parameters are in
square brackets. Presets are in parentheses.
? help, displays command list
ascii mode ascii, appends CR-LF per line
autoadapt toggles backoff function (on)
baudrate baud baud rate (16000 bits/sec)
binary mode binary, image transfer
blocksize bytes packet size in bytes (512 bytes/block)
bufferblock blks buffer size in blocks (128 blocks/buff)
burstsize packets burst size in packets (8 blocks/burst)
Polites, Wollman & Woo Experimental [Page 12]
RFC 1986 ETFTP August 1996
connect host [p] establish connection with host at port p
exit ends program
get rfile lfile copy remote file to local file
help same as ?
mode choice set transfer mode (binary)
multibuff num number of buffers (2 buffers)
put lfile rfile copy local file to remote file
quit same as exit
radiodelay sec transmission delay in seconds (2 sec)
status display network parameters
trace toggles debug display (off).
This is the scenario between client and server transfers:
Client sends OPEN for connection, blocksize, buffersize, burstsize,
burstrate, transfer mode, and get or put. See M bit of reserved
field.
Server sends a RESPONSE with the agreed parameters.
Receiver sends a CONTROL_GO request sending of first buffer.
Sender starts transfer by reading the file into multiple memory
buffers. See Figure 1, "File Segmentation,". Each buffer is divided
according to the number of bytes/block. Each block becomes a DATA
packet, which is concatenated according to the blocks/burst. Bursts
are transmitted according to the burstrate. Last data block is
flagged as LDATA type.
+---+ +---+ +---+ +---+ +---+ +---+ +---+ +---+
| | | 0 | | L | | 4 | | 3 | ---- | 2 | | 1 | | 0 |
| | | +---+ +---+ +---+ +---+ +---+ +---+ +---+
| | +-| | --> +---+ +---+ +---+ +---+ +---+
| | --> | 1 | | L | | 3 | ---- | 2 | | 1 | | 0 |
+---+ +---+ +---+ +---+ +---+ +---+ +---+
File Multi Buffers Blocks per Burst
Figure 1. File Segmentation
Receiver acknowledges buffer as CONTROL_OK or CONTROL_RESEND.
If blocks are missing, a CONTROL_RESEND packet is transmitted. If
half or more of the blocks in a buffer are missing, an adaptive
algorithm is used for the next buffer transfer. If no blocks are
missing, a CONTROL_OK packet is transmitted.
Polites, Wollman & Woo Experimental [Page 13]
RFC 1986 ETFTP August 1996
Sender re-transmits blocks until receipt of a CONTROL_OK. If the
adaptive algorithm is set, then new parameters are offered, in the
CONTROL_OK message. The priority of the adaptive algorithm is:
- Reduce packetsize by half (MIN = 16 bytes/packet)
- Reduce burstsize by one (MIN = 1 packet/burst)
- Reduce burstrate to actual tighttimer rate
If new parameters are valid, the sender transmits a NULL-ACK packet,
to confirm the changes.
Receiver sends a CONTROL_GO to request sending next buffer.
At end of transfer, sender sends a QUIT to close the connection.
Receiver acknowledges the close request with a DONE packet.
These parameters directly affect the throughput rate of ETFTP.
Packetsize The packetsize is the number of 8 bit bytes per
packet. This number refers to the user data bytes in a block, (frame),
exclusive of any network overhead. The packet size has a valid range
from 16 to 1,448 bytes. The Maximum Transfer Unit (MTU) implemented in
most commercial network devices is 1,500 bytes. The de-facto industry
standard is 576 byte packets.
Bufferblock The bufferblock is the number of blocks per buffer.
Each implementation may have restrictions on available memory, so the
buffersize is calculated by multiplying the packetsize times the
bufferblocks.
Baudrate The baudrate is the bits per second transfer rate of
the slowest link (i.e., the radios). The baudrate sets the speed of
the sending process. The sending process cannot detect the actual
speed of the network, so the user must set the correct baudrate.
Burstsize The burstsize in packets per burst sets how many
packets are concatenated and burst for transmission efficiency. The
burstsize times the packetsize must not exceed the available memory of
any intervening network devices. On the Ethernet portion of the
network, all the packets are sent almost instantaneously. It is
necessary to wait for the network to drain down its memory buffers,
before the next burst is sent. The sending process needs to regulate
the rate used to place packets into the network.
Polites, Wollman & Woo Experimental [Page 14]
RFC 1986 ETFTP August 1996
Radiodelay The radiodelay is the time in seconds per burst it
takes to synchronize with the radio controllers. Any additional
hardware delays should be set here. It is the aggregate delay of the
link layer, such as transmitter key-up, FEC, crypto synchronization,
and propagation delays.
These parameters above are used to calculate a burstrate, which is the
length of time it takes to transmit one burst. The ov is the overhead
of 72 bytes per packet of network encapsulation. A byte is defined as
8 bits. The burstrate value is:
burstrate = (packetsize+ov)*burstsize*8/baudrate
In a effort to calculate the round trip time, when data is flowing in
one direction for most of the transfer, the OPEN and RESPONSE message
types are timed, and the tactical radio delays are estimated. Using
only one packet in each direction to estimate the rate of the link is
statistically inaccurate. It was decided that the radio delay should
be a constant provided by the user interface. However, a default
value of 2 seconds is used. The granularity of this value is in
seconds because of two reasons. The first reason is that the UNIX
supports a sleep function in seconds only. The second reason is that
in certain applications, such as deep space probes, a 16-bit integer
maximum of 32,767 seconds would suffice.
From these parameters, several timers are derived. The control timer
is responsible for measuring the per buffer rate of transfer. The
SENDER copy is nicknamed the loosetimer.
loosetimer = (burstrate+radiodelay)*bufferblock/burstsize
The RECEIVER copy of the timer is nicknamed the tighttimer, which
measures the elapsed time between CONTROL_GO and CONTROL_OK packets.
The tighttimer is returned to the SENDER to allow the protocol to
adjust for the speed of the network.
The retransmit timer is responsible for measuring the network receive
data function. It is used to set an alarm signal (SIGALRM) to
interrupt the network read. The retransmit timer (wait) is initially
set to be the greater of twice the round trip or 4 times the
radiodelay, plus a constant 5 seconds.
Polites, Wollman & Woo Experimental [Page 15]
RFC 1986 ETFTP August 1996
wait = MAX ( 2*roundtriptime, 4*radiodelay ) + 5 seconds
and
alarm timeout = wait.
Each time the same read times out, a five second backoff is added to
the next wait. The backoff is necessary because the initial user
supplied radiodelay, or the initial measured round trip time may be
incorrect.
The retransmit timer is set differently for the RECEIVER during a
buffer transfer. Before the arrival of the first DATA packet, the
original alarm time out is used. Once the DATA packets start
arriving, and for the duration of each buffer transfer, the RECEIVER
alarm time out is reset to the expected arrival time of the last DATA
packet (blockstogo) plus the delay (wait). As each DATA packet is
received, the alarm is decremented by one packet interval. This same
algorithm is used for receiving missing packets, during a RESEND.
alarmtimeout = blockstogo*burstrate/burstsize + wait
The death timer is responsible for measuring the idle time of a
connection. In the ETFTP program, the death timer is set to be equal
to the accumulated time of ten re-transmissions plus their associated
backoffs. As such, the death timer value in the OPEN and RESPONSE
message types is un-necessary. In the ETFTP program, this field could
be used to transfer the radio delay value instead of creating the two
new fields.
The keepalive timer is responsible for resetting the death timer.
This timer will trigger the sending of a KEEPALIVE packet to prevent
the remote host from closing a connection due to the expiration of
its death timer. Due to the nature of the ETFTP server process, a
keepalive timer was not necessary, although it is implemented.
The NETBLT protocol has been tested on other high speed networks
before, see RFC 1030 [6]. These test results in Tables 13 and 14,
"ETFTP Performance," were gathered from files transferred across the
network and LST-5C TACSAT radios. The radios were connected together
via a coaxial cable to provide a "clean" link. A clean link is
defined to a BER of 10e-5. The throughput rates are defined to be the
file size divided by the elapsed time resulting in bits per second
(bps). The elapsed time is measured from the time of the "get" or
"put" command to the completion of the transfer. This is an all
inclusive time measurement based on user perspective. It includes the
Polites, Wollman & Woo Experimental [Page 16]
RFC 1986 ETFTP August 1996
connection time, transfer time, error recovery time, and disconnect
time. The user concept of elapsed time is the length of time it takes
to copy a file from disk to disk. These results show only the average
performances, including the occasional packet re-transmissions. The
network configuration was set as:
ETFTP Parameters:
Filesize 101,306 bytes
Radiodelay 2 seconds
Buffersize 16,384-131,072 bytes
Packetsize 512-2048 bytes
Burstsize 8-16 packets/burst
Gracilis PackeTen Parameters:
0 TX Delay 400 milliseconds
1 P Persist 255 [range 1-255]
2 Slot Time 30 milliseconds
3 TX Tail 300 milliseconds
4 Rcv Buffers 8 2048 bytes/buffer
5 Idle Code Flag
Radio Parameters:
Baudrate 16,000 bps
Encryption on
Table 13: ETFTP Performance at 8 Packets/Burst in Bits/Second
+-----------+-----------+-----------+-----------+-----------+
|buffersize |packetsize |packetsize |packetsize |packetsize |
|(bytes) |2,048 bytes|1,448 bytes|1,024 bytes|512 bytes |
+-----------+-----------+-----------+-----------+-----------+
| 16,384 | 7,153 | 6,952 | 6,648 | 5,248 |
+-----------+-----------+-----------+-----------+-----------+
| 32,768 | 7,652 | 7,438 | 7,152 | 4,926 |
+-----------+-----------+-----------+-----------+-----------+
| 65,536 | 8,072 | 8,752 | 8,416 | 5,368 |
+-----------+-----------+-----------+-----------+-----------+
| 131,072 | 8,828 | 9,112 | 7,888 | 5,728 |
+-----------+-----------+-----------+-----------+-----------+
Polites, Wollman & Woo Experimental [Page 17]
RFC 1986 ETFTP August 1996
Table 14: ETFTP Performance at 16 Packets/Burst in Bits/Second
+-----------+-----------+-----------+-----------+-----------+
|buffersize |packetsize |packetsize |packetsize |packetsize |
|(bytes) |2,048 bytes|1,448 bytes|1,024 bytes|512 bytes |
+-----------+-----------+-----------+-----------+-----------+
| 16,384 | 5,544 | 5,045 | 4,801 | 4,570 |
+-----------+-----------+-----------+-----------+-----------+
| 32,768 | 8,861 | 8,230 | 8,016 | 7,645 |
+-----------+-----------+-----------+-----------+-----------+
| 65,536 | 9,672 | 9,424 | 9,376 | 8,920 |
+-----------+-----------+-----------+-----------+-----------+
| 131,072 | 10,432 | 10,168 | 9,578 | 9,124 |
+-----------+-----------+-----------+-----------+-----------+
These tests were performed across a tactical radio link with a
maximum data rate of 16000 bps. In testing ETFTP, it was found that
the delay associated with the half duplex channel turnaround time was
the biggest factor in throughput performance. Therefore, every
attempt was made to minimize the number of times the channel needed
to be turned around. Obviously, the easiest thing to do is to use as
big a buffer as necessary to read in a file, as acknowledgments
occurred only at the buffer boundaries. This is not always feasible,
as available storage on disk could easily exceed available memory.
However, the current ETFTP buffersize is set at a maximum of 524,288
bytes.
The larger packetsizes also improved performance. The limit on
packetsize is based on the 1500 byte MTU of network store and forward
devices. In a high BER environment, a large packetsize could be
detrimental to success. By reducing the packetsize, even though it
negatively impacts performance, reliability is sustained. When used
in conjunction with FEC, both performance and reliability can be
maintained at an acceptable level.
The burstsize translates into how long the radio transmitters are
keyed to transmit. In ETFTP, the ideal situation is to have the first
packet of a burst arrive in the radio transmit buffer, as the last
packet of the previous burst is just finished being sent. In this
way, the radio transmitter would never be dropped for the duration of
one buffer. In a multi-user radio network, a full buffer transmission
would be inconsiderate, as the transmit cycle could last for several
minutes, instead of seconds. In measuring voice communications,
typical transmit durations are on the order of five to twenty
seconds. This means that the buffersize and burstsize could be
adjusted to have similar transmission durations.
Polites, Wollman & Woo Experimental [Page 18]
RFC 1986 ETFTP August 1996
[1] Clark, D., Lambert, M., and L. Zhang,
"NETBLT: A Bulk Data Transfer Protocol", RFC 998, MIT,
March 1987.
[2] Postel, J., "User Datagram Protocol" STD 6, RFC 768,
USC/Information Sciences Institute, August 1980.
[3] Sollins, K., "Trivial File Transfer Protocol", STD 33,
RFC 1350, MIT, July 1992.
[4] MIL-STD-2045-44500, 18 June 1993, "Military Standard Tactical
Communications Protocol 2 (TACO 2) fot the National Imagery
Transmission Format Standard", Ft. Monmouth, New Jersey.
[5] Stevens, W. Richard, 1990, "UNIX Network Programming",
Prentice-Hall Inc., Englewood, New Jersey, Chapter 12.
[6] Lambert, M., "On Testing the NETBLT Protocol over
Divers Networks", RFC 1030, MIT, November 1987.
The ETFTP program is a security loophole in any UNIX environment.
There is no user/password validation. All the problems associated to
TFTP are repeated in ETFTP. The server program must be owned by root
and setuid to root in order to work. As an experimental prototype
program, the security issue was overlooked. Since this protocol has
proven too be a viable solution in tactical radio networks, the
security issues will have to be addressed, and corrected.
Polites, Wollman & Woo Experimental [Page 19]
RFC 1986 ETFTP August 1996
ATD Advanced Technology Demonstration
AX.25 Amateur Radio X.25 Protocol
BER Bit Error Rate
EPLRS Enhanced Position Location Reporting Systems
ETFTP Enhanced Trivial File Transfer Protocol
FEC Forward Error Correction
FTP File Transfer Protocol
HF High Frequency
LCU Lightweight Computer Unit
ms milliseconds
MTU Maximum Transfer Unit
NETBLT NETwork Block Transfer protocol
NITFS National Imagery Transmission Format Standard
PC Personal Computer
RNC Radio Network Controller
SAS Survivable Adaptive Systems
SATCOM SATellite COMmunications
SCO Santa Cruz Operations
SINCGARS SINgle Channel Ground and Airborne Radio Systems
SLIP Serial Line Internet Protocol
TACO2 Tactical Communications Protocol 2
TCP Transmission Control Protocol
TFTP Trivial File Transfer Protocol
UDP User Datagram Protocol
UHF Ultra High Frequency
* Modification from NETBLT RFC 998.
* The new packet size is a modification to the NETBLT RFC 998.
* The new packet size is a modification to the NETBLT RFC 998.
Polites, Wollman & Woo Experimental [Page 21]