TFTP Blocksize Option
Let client and server negotiate larger blocks for faster file transfers.
Malkin, G. and A. Harkin, "TFTP Blocksize Option", RFC 1783, DOI 10.17487/RFC1783, March 1995, <https://www.rfc-editor.org/info/rfc1783>.
@misc{rfc1783,
series = {Request for Comments},
number = 1783,
howpublished = {RFC 1783},
publisher = {RFC Editor},
doi = {10.17487/RFC1783},
url = {https://www.rfc-editor.org/info/rfc1783},
author = {G. Malkin and A. Harkin},
title = {{TFTP Blocksize Option}},
pagetotal = 5,
year = 1995,
month = mar,
} RFC text
Contents
Network Working Group G. Malkin
Request for Comments: 1783 Xylogics, Inc.
Updates: 1350 A. Harkin
Category: Standards Track Hewlett Packard Co.
March 1995
TFTP Blocksize Option
Status of this Memo
This document specifies an Internet standards track protocol for the
Internet community, and requests discussion and suggestions for
improvements. Please refer to the current edition of the "Internet
Official Protocol Standards" (STD 1) for the standardization state
and status of this protocol. Distribution of this memo is unlimited.
Abstract
The Trivial File Transfer Protocol [1] is a simple, lock-step, file
transfer protocol which allows a client to get or put a file onto a
remote host. One of its primary uses is the booting of diskless
nodes on a Local Area Network. TFTP is used because it is very
simple to implement in a small node's limited ROM space. However,
the choice of a 512-byte blocksize is not the most efficient for use
on a LAN whose MTU may 1500 bytes or greater.
This document describes a TFTP option which allows the client and
server to negotiate a blocksize more applicable to the network
medium. The TFTP Option Extension mechanism is described in [2].
Blocksize Option Specification
The TFTP Read Request or Write Request packet is modified to include
the blocksize option as follows:
+-------+---~~---+---+---~~---+---+---~~---+---+---~~---+---+
| opc |filename| 0 | mode | 0 | blksize| 0 | #octets| 0 |
+-------+---~~---+---+---~~---+---+---~~---+---+---~~---+---+
opc
The opcode field contains either a 1, for Read Requests, or 2,
for Write Requests, as defined in [1].
filename
The name of the file to be read or written, as defined in [1].
This is a NULL-terminated field.
mode
The mode of the file transfer: "netascii", "octet", or "mail",
as defined in [1]. This is a NULL-terminated field.
blksize
The Blocksize option, "blksize" (case insensitive). This is a
NULL-terminated field.
#octets
The number of octets in a block, specified in ASCII. Valid
values range between "8" and "65464" octets, inclusive. This
is a NULL-terminated field.
For example:
+-------+--------+---+--------+---+--------+---+--------+---+
| 1 | foobar | 0 | binary | 0 | blksize| 0 | 1432 | 0 |
+-------+--------+---+--------+---+--------+---+--------+---+
is a Read Request, for the file named "foobar", in binary transfer
mode, with a block size of 1432 bytes (Ethernet MTU, less the UDP and
IP header lengths).
If the server is willing to accept the blocksize option, it sends an
Option Acknowledgment (OACK) to the client. The specified value must
be less than or equal to the value specified by the client. The
client must then either use the size specified in the OACK, or send
an ERROR packet, with error code 8, to terminate the transfer.
The rules for determining the final packet are unchanged from [1].
The reception of a data packet with a data length less than the
negotiated blocksize is the final packet. If the blocksize is
greater than the size of the packet, the first packet is the final
packet. If amount of data to be transfered is an integral multiple
of the blocksize, an extra data packet containing no data is sent to
end the transfer.
Proof of Concept
Performance tests were run on the prototype implementation using a
variety of block sizes. The tests were run on a lightly loaded
Ethernet, between two HP-UX 9000, in "octet" mode, on 2.25MB files.
The average (5x) transfer times for paths with (g-time) and without
(n-time) a intermediate gateway are graphed as follows:
|
37 + g
|
35 +
|
33 +
|
31 +
|
29 +
|
27 +
| g blocksize n-time g-time
25 + --------- ------ ------
s | n 512 23.85 37.05
e 23 + g 1024 16.15 25.65
c | 1432 13.70 23.10
o 21 + 2048 10.90 16.90
n | 4096 6.85 9.65
d 19 + 8192 4.90 6.15
s |
17 + g
| n
15 +
| n
13 +
|
11 + n
| g
9 +
|
7 + n
| g
5 + n
"
0 +------+------+--+---+------+------+---
512 1K | 2K 4K 8K
1432
blocksize (bytes)
The comparisons between transfer times (without a gateway) between
the standard 512-byte blocksize and the negotiated blocksizes are:
1024 2x -32%
1432 2.8x -42%
2048 4x -54%
4096 8x -71%
8192 16x -80%
As was anticipated, the transfer time decreases with an increase in
blocksize. The reason for the reduction in time is the reduction in
the number of packets sent. For example, by increasing the blocksize
from 512 bytes to 1024 bytes, not only are the number of data packets
halved, but the number of acknowledgement packets is also halved
(along with the number of times the data transmitter must wait for an
ACK). A secondary effect is the efficiency gained by reducing the
per-packet framing and processing overhead.
Of course, if the blocksize exceeds the path MTU, IP fragmentation
and reassembly will begin to add more overhead. This will be more
noticable the greater the number of gateways in the path.
Security Considerations
Security issues are not discussed in this memo.
References
[1] Sollins, K., "The TFTP Protocol (Revision 2)", STD 33, RFC 1350,
MIT, July 1992.
[2] Malkin, G., and A. Harkin, "TFTP Option Extension", RFC 1782,
Xylogics, Inc., Hewlett Packard Co., March 1995.
Authors' Addresses
Gary Scott Malkin
Xylogics, Inc.
53 Third Avenue
Burlington, MA 01803
Phone: (617) 272-8140
EMail: gmalkin@xylogics.com
Art Harkin
Internet Services Project
Information Networks Division
19420 Homestead Road MS 43LN
Cupertino, CA 95014
Phone: (408) 447-3755
EMail: ash@cup.hp.com
AI ELI5
On this page
§1TFTP and the Blocksize Problem
The Trivial File Transfer Protocol (TFTP) is a simple, lock-step protocol for transferring files. It was originally designed for booting diskless computers from ROM and uses 512-byte blocks, a size chosen for simplicity on resource-constrained machines.
Modern Ethernet and other LANs have MTU (Maximum Transmission Unit) sizes of 1500 bytes or larger. The original 512-byte blocksize wastes bandwidth because each block requires an acknowledgment packet, doubling the number of round trips. This document defines a TFTP option to negotiate a larger blocksize better suited to the network medium.
TFTP was designed for tiny machines and uses tiny 512-byte blocks. Modern networks can carry much bigger pieces, so why not send 4000 bytes at a time instead?
§2The Blocksize Option
The blocksize option is added to the TFTP Read or Write Request packet. The client includes the option name "blksize" and a proposed size (8 to 65464 octets) as null-terminated ASCII strings. For example, a 1432-byte blocksize matches Ethernet MTU minus UDP and IP header lengths.
If the server accepts the option, it sends an Option Acknowledgment (OACK) with a blocksize value less than or equal to the client's proposal. The client must then either use the agreed size or send an ERROR packet to abort.
The client asks for bigger blocks in its initial request. The server agrees by sending back the size it will use, which is no bigger than what the client asked for.
§3Transfer Speed Improvement
Performance testing on Ethernet with 2.25 MB files showed dramatic improvements. With a 512-byte blocksize, transfers took 23.85 seconds. Increasing to 1432 bytes reduced time to 13.70 seconds (42% faster). A 4096-byte blocksize achieved 6.85 seconds (71% faster), and 8192 bytes reached 4.90 seconds (80% faster).
The speedup comes from fewer packets and acknowledgments. Doubling the blocksize roughly halves the number of packets, and the secondary effect is reduced per-packet framing overhead. However, if the blocksize exceeds the path MTU, IP fragmentation will add overhead, more noticeable with multiple gateways.
Bigger blocks mean fewer packets, fewer waits for acknowledgments, and less overhead. Tests showed 8x faster transfers with very large blocks.