UNIX Network Software Status

This paper describes the present state of the software.  The kernel and user
code are discussed separately.

Kernel:
	Driver for "Pronet" (10Mb/s token-passing ring net) installed
	and working.
	Internet packet transmission installed and working.
	Internet packet reception, and demultiplexing of UDP and TCP
	packets, installed and working.  Other protocols are received
	but not demultiplexed by the kernel.
	Reassembly of fragmented internet packets installed and working.
	Maintains cache of internet host ids with the best local gateway
	to which packets destined for each host should be sent; this
	cache is updated by calls from user programs.
	Demultiplexing of incoming ICMP error packets written and tested,
	but not yet installed (again, no one sends us ICMP errors).
	Pseudo-teletype code installed and working (not really net software,
	but necessary for Server Telnet).

User:
	ICMP/GGP daemon installed and working.  It presently handles only
	GGP and ICMP echo packets, ICMP timestamp packets, and ICMP redirects.
	It also generates GGP and ICMP echoes and timestamps on user request.
	The code to process incoming redirects and update the kernel
	routing tables is working.

	User and server TFTP daemons written and working. These are our primary
	file transfer programs at present.

	User TCP/Telnet program written and working.  This is based on a
	"toy" TCP which works, but provides minimal functionality.  It in
	turn is based on a small multitasking package which permits multiple
	tasks (each with its own stack) to run within a single UNIX process.

	User and server TCP/SMTP programs written and working.  These also
	use the tasking-based TCP.  They are integrated into our mail system,
	which is homegrown, and hence will have to be modified for use with
	anyone else's mailer.

	Server TCP/Telnet written and working.  The TCP is believed to meet
	the full specification, including handling the maximum buffer size
	option, and the security and precedence fields.  Server Telnet
	requires the installation of pseudo-teletypes in the kernel.

Libraries:
	We presently have libraries of routines to perform the basic Internet
	and UDP functions: connection open/close, packet allocation/
	deallocation, and packet transmission/reception.  We also have
	library routines for host name-to-internet address resolution,
	by querying UDP nameservers.  The internet library code performs
	packet fragmentation; it permits the user to specify an option
	string in outgoing packets and to obtain the option string from
	incoming packets, but performs no option processing itself.

Documentation:
	Documentation is unfortunately miminal at this point.  We have a
	short document which gives an overview of the kernel code, describes
	the kernel calls, and explains some of the rationale behind the
	design.  We have manual pages for the Internet and UDP libraries.
	We have a short paper on the performance testing we have done
	so far.  There is virtually no documentation on the ICMP or
	TFTP daemons, user Telnet, toy TCP, or tasking.  Some of this
	lack of documentation should be remedied soon.

We have the following (incomplete) performance figures.  All measurements
were made on an unloaded pdp-11/45 with 8K cache, 3/4 Mbyte of main memory,
6 512-byte kernel packet buffers, and a Pronet net interface. Transmission 
of a 512-byte packet takes on average 7 milliseconds; reception takes on 
average 4 milliseconds (reception time does not include process scheduling 
time).  Round-trip time for an ICMP timestamp packet sent from our machine 
to itself (thus requiring 2 packet transmissions and 2 receptions) takes 
between 30 and 40 milliseconds. The maximum TFTP transfer rate we have observed
was 133 Kbits/second, between our machine and a VAX on the same ring net (the 
maximum transfer rate of our disks seems to be about 140 Kbits/sec).  The toy 
TCP used in the user Telnet has been observed to sink data (ie memory-to-memory
transfer) from the VAX at 215 Kbits/second.  The TCP used in the server Telnet 
has been observed to source data (memory-to-memory transfer) to an Alto on an
ethernet at 315 Kbits/second; the gateway between Pronet and ethernet was an 
LSI 11/03.

Finally, a few notes on what our implementation does NOT support.
	1) We do not support more than one network interface per
	   host.
	2) The only network driver we have written is for the Pronet
	   ring.  We do not have and do not plan an IMP driver.
	3) Work on FTP is in a preliminary stage; we have no estimate
	   of when it will be done.
	4) This is not a "polished" distribution, nor is it supplied
	   ready-to-run.  Some work will be required to bring things
	   up.