Message-ID: < anews.Aucbvax.5236>
Date: Wed Nov 18 05:50:19 1981
Subject: TCP-IP Digest, Vol 1 #6
>From tcp-ip@brl Wed Nov 18 05:33:31 1981
TCP/IP Digest Wednesday, 11 Nov 1981 Volume 1 : Issue 6
Administrivia: Long Delay
Disabling NCPs -- Directory Service?
TOPS-20 TCP Defenses? -- TCP/IP for Cybers?
Unix/ArpaNet TCP Problems & Solutions
TCP/IP Performance on VAXen
Berkeley Enhanced TCP/IP for 4BSD UNIX
From: Mike Muuss
Hello again folks! Sorry about the long delay between this digest and
the previous one -- I was in Denver on travel, and could not get to a
terminal. I hope to send out future digests as soon as enough material
has been accumulated.
From: POSTEL at USC-ISIF
Subject: Disabling NCPs
There has been some talk of "forcing" the move to TCP by various
administrative and policy measures. There was also a claim that
there was no technical way to force the abandonment of NCP. It
should be pointed out that a quite simple modification to the IMP
program would enable the IMPs to filter out and discard all NCP
traffic. As far as i know, there has been no decision to do this,
but you should be aware that it is technical feasible.
From: Harris A. Meyers < Meyers at SRI-KL>
A useful fuction for this list to perform might be to produce a directory of
who has IP/TCP available for what machines and operating systems.
In particular I am looking for a version to run on IBM's VM/CMS.
Subject: Re: TOPS 20 TCP/IP
From: mike at RAND-UNIX
I have often heard criticisms of TOPS 20 TCP/IP implementation, but
never a defense. Does anyone from BBN or ARPA care to defend their
implementation or do they agree with the criticisms?
From: roya at UTEXAS-11
Subject: Is there TCP/IP for Cyber machines?
Do any of you know any sites that might have or is planning to
implement TCP/IP for CDC Cyber machines like 175 with NOS like
[ Tektronix Labs (Clem Cole et.al.) have already implemented
TCP and IP in RatFor for their Cyber, running under NOS. -Mike ]
From: greg at NPRDC
Subject: Future contributions
Now that all the special interest groups have spoken, can we get back to
the original subject? In case you've forgotten, it was "Unix/ARPAnet TCP
problems and solutions." Although I'm interested in the various problems/
possibilities of using TCP on other operating systems or other ethers, at
a minimum, our mutual interest is getting our machines running TCP before
the deadline. (Probabally this list goes a little farther than that; to
those people, I appologize. But we are the ones with the deadline fast
approaching.) Maybe we can discuss theoretical issues later, but I am
more interested in the practical issues -- namely, who has TCP up? How
is it connected to the ARPAnet (or even another ether, if the problems/
solutions are similar)? What problems were encountered? How fast is it?
How does it compare in simplicity/performance/transparancy/completeness/
functionality/limitations/etc. with the other possibilities? So far, we
have heard of two real choices (assuming that we're not going to have to
buy any additional hardware): BBN and 3COM. Who's got them up? How
connected? (I am VDH, so solutions that don't have a VDH driver are
uninteresting.) Speak up; now's your chance to brag, and you can do the
rest of us a real service.
[ Actually, I had hoped that this digest could serve as a forum for
technical discussion of networking for ALL systems, but clearly the
transition to TCP for current ArpaNet Hosts is the primary motivator.
I hope that this list will not restrict itself just to UNIX, though.
Subject: TCP/IP Performance on VAX
From: CERF at USC-ISI
UC Berkeley has been developing a paging UNIX(TM) for the VAX
based on V7 UNIX (TM-Western Electric). BBN has been developing
a TCP/IP for this VAX UNIX(TM) and UCB recently reported data
bandwidths of Mb/sec over a 3 Mb/s Ethernet running TCP/IP to
TCP/IP including checksumming. This figure obtained on VAX
11/750 using 1 kilobyte packets.
A point of contact for the Berkeley + BBN VAX UNIX(TM) is Prof.
Robert Fabry (ARPAVAX.Fabry@Berkeley).
A point of contact at BBN is Gurwitz@BBN-RSM(Rob Gurwitz).
Vint Cerf (bd)
From: ARPAVAX .wnj at Berkeley
Subject: tcp-ip digest contribution
Moderator: Here is a description of the work we are doing a
Berkeley with tcp-ip. I hope it is in time for this weeks
digest. We have enjoyed the past digests and hope that future
digests will be as interesting.
===== begin =====
The Computer Systems Research Group at Berkeley is
enhancing the UNIX operating system with DARPA support.
We are improving UNIX memory management facilities, working
on extensions to UNIX to support better inter-process communication, and
incorporating support for both local and long haul networks. In
particular, we expect to try using the INTERNET protocols on a number
of different commercially available local network interfaces.
The basis for our INTERNET protocols is the TCP/IP
implementation done by Rob Gurwitz at BBN. While this TCP has more
than adequate performance for use in the ARPANET context, we need
extremely good performance to be able to use the protocols as the basis
for construction of distributed UNIX applications between machines on a
local network. In particular, we wish to insure that data can be
transferred rapidly between VAX machines on 10 Megabaud Ethernet
cables. Our current file system organization uses 1024 byte records,
while a newer, high-performance file system will use 4096 byte
records. We are therefore interested in the effective transmission of
records of these sizes. We are also interested in low per-packet
overhead for distributed applications which exchange many messages.
We have just finished about three weeks of tuning of the BBN TCP/IP
for our 3 Megabaud prototype Ethernet. We had previously brought
TCP/IP up on the Ethernet and were interested in learning more about
the internals of TCP and discovering whether the protocol would be a
bottleneck when running on a local network. The results we have
obtained suggest that this is not the case.
As an experiment to investigate the performance of the resulting
TCP/IP implementation, we transmitted 4 Megabytes of data between two
user processes on different machines. The transfer was partitioned
into 1024 byte records and encapsulated in 1068 byte Ethernet packets.
Sending the data from our 11/750 to our 11/780 through TCP/IP takes 28
seconds. This includes all the time needed to set up and tear down the
connections, for an user-user throughput of 1.2 Megabaud. During this
time the 11/750 is CPU saturated, but the 11/780 has about 30% idle
time. The time spent in the system processing the data is spread out
among handling for the Ethernet (20%), IP packet processing (10%), TCP
processing (30%), checksumming (25%), and user system call handling
(15%), with no single part of the handling dominating the time in the
The TCP performance exceeds the throughput of the current
VAX/UNIX file system by a factor of 3, due to the small block size of
the UNIX file system. It comes within a factor of 2 of our per-spindle
performance of a early prototype of a new file system organization we
are working on. The relative speed of the TCP/IP protocol and the file
system suggests that we will be able to transfer volumes of data
regularly between machines without any special protocols. The limited
bandwidth of our 3 Megabaud cable may be a bottleneck until we put in
one or more 10 Megabaud cables.
Higher rates can be expected between VAX-11/780's, but we have
no direct measurements yet, as we have only one 780 with easily accessible
down time. Improvements are yet possible within the bounds of the
TCP/IP protocol and the new Ethernet standard (which limits packets to
about 1500 bytes). In particular, we may use IP fragmentation and
reassembly on the local network to allow the TCP and higher level
system code to process 4096 byte data records (which is a more natural
block size for the newer system we are working on, being a basic file
system data page size.) This is convenient within the bounds of the
Ethernet standards only because IP supports fragmentation and
reassembly in a general way. Simple techniques are also available
which would reduce the number of ACK's by a significant amount to
further speed the TCP.
Using information gathered from UNIX kernel profiling we can
estimate the speed improvements possible given a 10 Megabaud cable.
(All of these projections will be for user-user throughput between
11/750's.) Switching to 4096 byte segments in TCP transfers, while
fragmenting at the IP layer (to stay within the packet length
restriction of the Ethernet standard) we estimate an increased
throughput of 1.9 Megabaud. This is with an interface which is
functionality equivalent to our current prototype Ethernets. We plan
on changing UNIX soon so that user i/o buffers are naturally page aligned.
This would eliminate a copy of the data (when it is read) to raise the
throughput to about 2.25 Megabaud. (This speedup only at the receiver
gives an end-end speedup because the receiving process otherwise has
more work to do.) With checksum calculation support from the interface
hardware the user-user bandwidth would rise to about 3.5 Megabaud.
At this point the major overhead is the processing of the 4 interrupts
for the fragments of the 4k packets; a transmission medium which
allowed us to use 4k packets would let the bandwidth rise to about
6.6 Megabaud, user-user. A final improvement would be to implement a
variant on the send system call which released the virtual memory when
the message was sent. This would be very useful for servers and could
also be used by the standard i/o library. This would reduce the
transmit overhead (which at this point should be greater than the
receive overhead) making the final throughput about 8 Megabaud.
On 11/780's, these numbers typically scale up by 1.4 so that we
can project the throughput with the improvements described above to be
about 11.2 Megabaud, user-user. While factors in the VAX architecture other
than the protocol might well dominate before this bandwidth is achieved,
this means that high data rates through the protocol can be achieved
with a relatively small percentage of the processor.
We are working on IPC facilities for UNIX which will interface
to the INTERNET protocol family, and allow us to construct distributed
applications for UNIX without explicit dependence on the network layer
protocols. The measurements reported here suggest that we do not need
to look for more efficient local network protocols, but will need
to support other protocols only for inter-operability with other
networks and systems.
We will be working with Rob Gurwitz at BBN in the coming weeks,
combining our version of TCP/IP with his current version. We look
forward to making a high-performance version of the protocol available
to the VAX/UNIX community at an early date.
Bill Joy and Bob Fabry
End of TCP-IP Digest