Network Working Group M. Lambert
Request for Comments: 1030 M.I.T. Laboratory for Computer Science
November 1987
On Testing the NETBLT Protocol over Divers Networks
STATUS OF THIS MEMO
This RFC describes the results gathered from testing NETBLT over
three networks of differing bandwidths and round-trip delays. While
the results are not complete, the information gathered so far has
been very promising and supports RFC-998's assertion that that NETBLT
can provide very high throughput over networks with very different
characteristics. Distribution of this memo is unlimited.
NETBLT (NETwork BLock Transfer) is a transport level protocol
intended for the rapid transfer of a large quantity of data between
computers. It provides a transfer that is reliable and flow
controlled, and is designed to provide maximum throughput over a wide
variety of networks. The NETBLT protocol is specified in RFC-998;
this document assumes an understanding of the specification as
described in RFC-998.
Tests over three different networks are described in this document.
The first network, a 10 megabit-per-second Proteon Token Ring, served
as a "reference environment" to determine NETBLT's best possible
performance. The second network, a 10 megabit-per-second Ethernet,
served as an access path to the third network, the 3 megabit-per-
second Wideband satellite network. Determining NETBLT's performance
over the Ethernet allowed us to account for Ethernet-caused behaviour
in NETBLT transfers that used the Wideband network. Test results for
each network are described in separate sections. The final section
presents some conclusions and further directions of research. The
document's appendices list test results in detail.
Many thanks are due Bob Braden, Stephen Casner, and Annette DeSchon
of ISI for the time they spent analyzing and commenting on test
results gathered at the ISI end of the NETBLT Wideband network tests.
Bob Braden was also responsible for porting the IBM PC/AT NETBLT
implementation to a SUN-3 workstation running UNIX. Thanks are also
due Mike Brescia, Steven Storch, Claudio Topolcic and others at BBN
who provided much useful information about the Wideband network, and
M. Lambert [Page 1]
RFC 1030 Testing the NETBLT Protocol November 1987
helped monitor it during testing.
This section briefly describes the NETBLT implementations and test
programs used in the testing. Currently, NETBLT runs on three
machine types: Symbolics LISP machines, IBM PC/ATs, and SUN-3s. The
test results described in this paper were gathered using the IBM
PC/AT and SUN-3 NETBLT implementations. The IBM and SUN
implementations are very similar; most differences lie in timer and
multi-tasking library implementations. The SUN NETBLT implementation
uses UNIX's user-accessible raw IP socket; it is not implemented in
the UNIX kernel.
The test application performs a simple memory-to-memory transfer of
an arbitrary amount of data. All data are actually allocated by the
application, given to the protocol layer, and copied into NETBLT
packets. The results are therefore fairly realistic and, with
appropriately large amounts of buffering, could be attained by disk-
based applications as well.
The test application provides several parameters that can be varied
to alter NETBLT's performance characteristics. The most important of
these parameters are:
burst interval The number of milliseconds from the start of one
burst transmission to the start of the next burst
transmission.
burst size The number of packets transmitted per burst.
buffer size The number of bytes in a NETBLT buffer (all
buffers must be the same size, save the last,
which can be any size required to complete the
transfer).
data packet size
The number of bytes contained in a NETBLT DATA
packet's data segment.
number of outstanding buffers
The number of buffers which can be in
transmission/error recovery at any given moment.
M. Lambert [Page 2]
RFC 1030 Testing the NETBLT Protocol November 1987
The protocol's throughput is measured in two ways. First, the "real
throughput" is throughput as viewed by the user: the number of bits
transferred divided by the time from program start to program finish.
Although this is a useful measurement from the user's point of view,
another throughput measurement is more useful for analyzing NETBLT's
performance. The "steady-state throughput" is the rate at which data
is transmitted as the transfer size approaches infinity. It does not
take into account connection setup time, and (more importantly), does
not take into account the time spent recovering from packet-loss
errors that occur after the last buffer in the transmission is sent
out. For NETBLT transfers using networks with long round-trip delays
(and consequently with large numbers of outstanding buffers), this
"late" recovery phase can add large amounts of time to the
transmission, time which does not reflect NETBLT's peak transmission
rate. The throughputs listed in the test cases that follow are all
steady-state throughputs.
This section describes the theoretical performance of the IBM PC/AT
NETBLT implementation on both the transmitting and receiving sides.
Theoretical performance was measured on two LANs: a 10 megabit-per-
second Proteon Token Ring and a 10 megabit-per-second Ethernet.
"Theoretical performance" is defined to be the performance achieved
if the sending NETBLT did nothing but transmit data packets, and the
receiving NETBLT did nothing but receive data packets.
Measuring the send-side's theoretical performance is fairly easy,
since the sending NETBLT does very little more than transmit packets
at a predetermined rate. There are few, if any, factors which can
influence the processing speed one way or another.
Using a Proteon P1300 interface on a Proteon Token Ring, the IBM
PC/AT NETBLT implementation can copy a maximum-sized packet (1990
bytes excluding protocol headers) from NETBLT buffer to NETBLT data
packet, format the packet header, and transmit the packet onto the
network in about 8 milliseconds. This translates to a maximum
theoretical throughput of 1.99 megabits per second.
Using a 3COM 3C500 interface on an Ethernet LAN, the same
implementation can transmit a maximum-sized packet (1438 bytes
excluding protocol headers) in 6.0 milliseconds, for a maximum
theoretical throughput of 1.92 megabits per second.
Measuring the receive-side's theoretical performance is more
difficult. Since all timer management and message ACK overhead is
incurred at the receiving NETBLT's end, the processing speed can be
slightly slower than the sending NETBLT's processing speed (this does
M. Lambert [Page 3]
RFC 1030 Testing the NETBLT Protocol November 1987
not even take into account the demultiplexing overhead that the
receiver incurs while matching packets with protocol handling
functions and connections). In fact, the amount by which the two
processing speeds differ is dependent on several factors, the most
important of which are: length of the NETBLT buffer list, the number
of data timers which may need to be set, and the number of control
messages which are ACKed by the data packet. Almost all of this
added overhead is directly related to the number of outstanding
buffers allowable during the transfer. The fewer the number of
outstanding buffers, the shorter the NETBLT buffer list, and the
faster a scan through the buffer list and the shorter the list of
unacknowledged control messages.
Assuming a single-outstanding-buffer transfer, the receiving-side
NETBLT can DMA a maximum-sized data packet from the Proteon Token
Ring into its network interface, copy it from the interface into a
packet buffer and finally copy the packet into the correct NETBLT
buffer in 8 milliseconds: the same speed as the sender of data.
Under the same conditions, the implementation can receive a maximum-
sized packet from the Ethernet in 6.1 milliseconds, for a maximum
theoretical throughput of 1.89 megabits per second.
The Proteon Token Ring used for testing is a 10 megabit-per-second
LAN supporting about 40 hosts. The machines on either end of the
transfer were IBM PC/ATs using Proteon P1300 network interfaces. The
Token Ring provides high bandwidth with low round-trip delay and
negligible packet loss, a good debugging environment in situations
where packet loss, packet reordering, and long round-trip time would
hinder debugging. Also contributing to high performance is the large
(maximum 2046 bytes) network MTU. The larger packets take somewhat
longer to transmit than do smaller packets (8 milliseconds per 2046
byte packet versus 6 milliseconds per 1500 byte packet), but the
lessened per-byte computational overhead increases throughput
somewhat.
The fastest single-outstanding-buffer transmission rate was 1.49
megabits per second, and was achieved using a test case with the
following parameters:
M. Lambert [Page 4]
RFC 1030 Testing the NETBLT Protocol November 1987
transfer size 2-5 million bytes
data packet size
1990 bytes
buffer size 19900 bytes
burst size 5 packets
burst interval 40 milliseconds. The timer code on the IBM PC/AT
is accurate to within 1 millisecond, so a 40
millisecond burst can be timed very accurately.
Allowing only one outstanding buffer reduced the protocol to running
"lock-step" (the receiver of data sends a GO, the sender sends data,
the receiver sends an OK, followed by a GO for the next buffer).
Since the lock-step test incurred one round-trip-delay's worth of
overhead per buffer (between transmission of a buffer's last data
packet and receipt of an OK for that buffer/GO for the next buffer),
a test with two outstanding buffers (providing essentially constant
packet transmission) should have resulted in higher throughput.
A second test, this time with two outstanding buffers, was performed,
with the above parameters identical save for an increased burst
interval of 43 milliseconds. The highest throughput recorded was
1.75 megabits per second. This represents 95% efficiency (5 1990-
byte packets every 43 milliseconds gives a maximum theoretical
throughput of 1.85 megabits per second). The increase in throughput
over a single-outstanding-buffer transmission occurs because, with
two outstanding buffers, there is no round-trip-delay lag between
buffer transmissions and the sending NETBLT can transmit constantly.
Because the P1300 interface can transmit and receive concurrently, no
packets were dropped due to collision on the interface.
As mentioned previously, the minimum transmission time for a
maximum-sized packet on the Proteon Ring is 8 milliseconds. One
would expect, therefore, that the maximum throughput for a double-
buffered transmission would occur with a burst interval of 8
milliseconds times 5 packets per burst, or 40 milliseconds. This
would allow the sender of data to transmit bursts with no "dead time"
in between bursts. Unfortunately, the sender of data must take time
to process incoming control messages, which typically forces a 2-3
millisecond gap between bursts, lowering the throughput. With a
burst interval of 43 milliseconds, the incoming packets are processed
M. Lambert [Page 5]
RFC 1030 Testing the NETBLT Protocol November 1987
during the 3 millisecond-per-burst "dead time", making the protocol
more efficient.
The network used in performing this series of tests was a 10 megabit
per second Ethernet supporting about 150 hosts. The machines at
either end of the NETBLT connection were IBM PC/ATs using 3COM 3C500
network interfaces. As with the Proteon Token Ring, the Ethernet
provides high bandwidth with low delay. Unfortunately, the
particular Ethernet used for testing (MIT's infamous Subnet 26) is
known for being somewhat noisy. In addition, the 3COM 3C500 Ethernet
interfaces are relatively unsophisticated, with only a single
hardware packet buffer for both transmitting and receiving packets.
This gives the interface an annoying tendency to drop packets under
heavy load. The combination of these factors made protocol
performance analysis somewhat more difficult than on the Proteon
Ring.
The fastest single-buffer transmission rate was 1.45 megabits per
second, and was achieved using a test case with the following
parameters:
transfer size 2-5 million bytes
data packet size
1438 bytes (maximum size excluding protocol
headers).
buffer size 14380 bytes
burst size 5 packets
burst interval 30 milliseconds (6.0 milliseconds x 5 packets).
A second test, this one with parameters identical to the first save
for number of outstanding buffers (2 instead of 1) resulted in
substantially lower throughput (994 kilobits per second), with a
large number of packets retransmitted (10%). The retransmissions
occurred because the 3COM 3C500 network interface has only one
hardware packet buffer and cannot hold a transmitting and receiving
packet at the same time. With two outstanding buffers, the sender of
data can transmit constantly; this means that when the receiver of
data attempts to send a packet, its interface's receive hardware goes
M. Lambert [Page 6]
RFC 1030 Testing the NETBLT Protocol November 1987
deaf to the network and any packets being transmitted at the time by
the sender of data are lost. A symmetrical problem occurs with
control messages sent from receiver of data to sender of data, but
the number of control messages sent is small enough and the
retransmission algorithm redundant enough that little performance
degradation occurs due to control message loss.
When the burst interval was lengthened from 30 milliseconds per 5
packet burst to 45 milliseconds per 5 packet burst, a third as many
packets were dropped, and throughput climbed accordingly, to 1.12
megabits per second. Presumably, the longer burst interval allowed
more dead time between bursts and less likelihood of the receiver of
data's interface being deaf to the net while the sender of data was
sending a packet. An interesting note is that, when the same test
was conducted on a special Ethernet LAN with the only two hosts
attached being the two NETBLT machines, no packets were dropped once
the burst interval rose above 40 milliseconds/5 packet burst. The
improved performance was doubtless due to the absence of extra
network traffic.
The following section describes results gathered using the Wideband
network. The Wideband network is a satellite-based network with ten
stations competing for a raw satellite channel bandwidth of 3
megabits per second. Since the various tests resulted in substantial
changes to the NETBLT specification and implementation, some of the
major changes are described along with the results and problems that
forced those changes.
The Wideband network has several characteristics that make it an
excellent environment for testing NETBLT. First, it has an extremely
long round-trip delay (1.8 seconds). This provides a good test of
NETBLT's rate control and multiple-buffering capabilities. NETBLT's
rate control allows the packet transmission rate to be regulated
independently of the maximum allowable amount of outstanding data,
providing flow control as well as very large "windows". NETBLT's
multiple-buffering capability enables data to still be transmitted
while earlier data are awaiting retransmission and subsequent data
are being prepared for transmission. On a network with a long
round-trip delay, the alternative "lock-step" approach would require
a 1.8 second gap between each buffer transmission, degrading
performance.
Another interesting characteristic of the Wideband network is its
throughput. Although its raw bandwidth is 3 megabits per second, at
the time of these tests fully 2/3 of that was consumed by low-level
network overhead and hardware limitations. (A detailed analysis of
M. Lambert [Page 7]
RFC 1030 Testing the NETBLT Protocol November 1987
the overhead appears at the end of this document.) This reduces the
available bandwidth to just over 1 megabit per second. Since the
NETBLT implementation can run substantially faster than that, testing
over the Wideband net allows us to measure NETBLT's ability to
utilize very high percentages of available bandwidth.
Finally, the Wideband net has some interesting packet reorder and
delay characteristics that provide a good test of NETBLT's ability to
deal with these problems.
Testing progressed in several phases. The first phase involved using
source-routed packets in a path from an IBM PC/AT on MIT's Subnet 26,
through a BBN Butterfly Gateway, over a T1 link to BBN, onto the
Wideband network, back down into a BBN Voice Funnel, and onto ISI's
Ethernet to another IBM PC/AT. Testing proceeded fairly slowly, due
to gateway software and source-routing bugs. Once a connection was
finally established, we recorded a best throughput of approximately
90K bits per second.
Several problems contributed to the low throughput. First, the
gateways at either end were forwarding packets onto their respective
LANs faster than the IBM PC/AT's could accept them (the 3COM 3C500
interface would not have time to re-enable input before another
packet would arrive from the gateway). Even with bursts of size 1,
spaced 6 milliseconds apart, the gateways would aggregate groups of
packets coming from the same satellite frame, and send them faster
than the PC could receive them. The obvious result was many dropped
packets, and degraded performance. Also, the half-duplex nature of
the 3COM interface caused incoming packets to be dropped when packets
were being sent.
The number of packets dropped on the sending NETBLT side due to the
long interface re-enable time was reduced by packing as many control
messages as possible into a single control packet (rather than
placing only one message in a control packet). This reduced the
number of control packets transmitted to one per buffer transmission,
which the PC was able to handle. In particular, messages of the form
OK(n) were combined with messages of the form GO(n + 1), in order to
prevent two control packets from arriving too close together to both
be received.
Performance degradation from dropped control packets was also
minimized by changing to a highly redundant control packet
transmission algorithm. Control messages are now stored in a single
long-lived packet, with ACKed messages continuously bumped off the
head of the packet and new messages added at the tail of the packet.
Every time a new message needs to be transmitted, any unACKed old
messages are transmitted as well. The sending NETBLT, which receives
M. Lambert [Page 8]
RFC 1030 Testing the NETBLT Protocol November 1987
these control messages, is tuned to ignore duplicate messages with
almost no overhead. This transmission redundancy puts little
reliance on the NETBLT control timer, further reducing performance
degradation from lost control packets.
Although the effect of dropped packets on the receiving NETBLT could
not be completely eliminated, it was reduced somewhat by some changes
to the implementation. Data packets from the sending NETBLT are
guaranteed to be transmitted by buffer number, lowest number first.
In some cases, this allowed the receiving NETBLT to make retransmit-
request decisions for a buffer N, if packets for N were expected but
none were received at the time packets for a buffer N+M were
received. This optimization was somewhat complicated, but improved
NETBLT's performance in the face of missing packets. Unfortunately,
the dropped-packet problem remained until the NETBLT implementation
was ported to a SUN-3 workstation. The SUN is able to handle the
incoming packets quite well, dropping only 0.5% of the data packets
(as opposed to the PC's 15 - 20%).
Another problem with the Wideband network was its tendency to re-
order and delay packets. Dealing with these problems required
several changes in the implementation. Previously, the NETBLT
implementation was "optimized" to generate retransmit requests as
soon as possible, if possible not relying on expiration of a data
timer. For instance, when the receiving NETBLT received an LDATA
packet for a buffer N, and other packets in buffer N had not arrived,
the receiver would immediately generate a RESEND for the missing
packets. Similarly, under certain circumstances, the receiver would
generate a RESEND for a buffer N if packets for N were expected and
had not arrived before packets for a buffer N+M. Obviously, packet-
reordering made these "optimizations" generate retransmit requests
unnecessarily. In the first case, the implementation was changed to
no longer generate a retransmit request on receipt of an LDATA with
other packets missing in the buffer. In the second case, a data
timer was set with an updated (and presumably more accurate) value,
hopefully allowing any re-ordered packets to arrive before timing out
and generating a retransmit request.
It is difficult to accommodate Wideband network packet delay in the
NETBLT implementation. Packet delays tend to occur in multiples of
600 milliseconds, due to the Wideband network's datagram reservation
scheme. A timer value calculation algorithm that used a fixed
variance on the order of 600 milliseconds would cause performance
degradation when packets were lost. On the other hand, short fixed
variance values would not react well to the long delays possible on
the Wideband net. Our solution has been to use an adaptive data
timer value calculation algorithm. The algorithm maintains an
average inter-packet arrival value, and uses that to determine the
M. Lambert [Page 9]
RFC 1030 Testing the NETBLT Protocol November 1987
data timer value. If the inter-packet arrival time increases, the
data timer value will lengthen.
At this point, testing proceeded between NETBLT implementations on a
SUN-3 workstation and an IBM PC/AT. The arrival of a Butterfly
Gateway at ISI eliminated the need for source-routed packets; some
performance improvement was also expected because the Butterfly
Gateway is optimized for IP datagram traffic.
In order to put the best Wideband network test results in context, a
short analysis follows, showing the best throughput expected on a
fully loaded channel. Again, a detailed analysis of the numbers that
follow appears at the end of this document.
The best possible datagram rate over the current Wideband
configuration is 24,054 bits per channel frame, or 3006 bytes every
21.22 milliseconds. Since the transmission route begins and ends on
an Ethernet, the largest amount of data transmissible (after
accounting for packet header overhead) is 1438 bytes per packet.
This translates to approximately 2 packets per frame. Since we want
to avoid overflowing the channel, we should transmit slightly slower
than the channel frame rate of 21.2 milliseconds. We therefore came
up with a best possible throughput of 2 1438-byte packets every 22
milliseconds, or 1.05 megabits per second.
Because of possible software bugs in either the Butterfly Gateway or
the BSAT (gateway-to-earth-station interface), 1438-byte packets were
fragmented before transmission over the Wideband network, causing
packet delay and poor performance. The best throughput was achieved
with the following values:
transfer size 500,000 - 750,000 bytes
data packet size
1432 bytes
buffer size 14320 bytes
burst size 5 packets
burst interval 55 milliseconds
Steady-state throughputs ranged from 926 kilobits per second to 942
kilobits per second, approximately 90% channel utilization. The
M. Lambert [Page 10]
RFC 1030 Testing the NETBLT Protocol November 1987
amount of data transmitted should have been an order of magnitude
higher, in order to get a longer steady-state period; unfortunately
at the time we were testing, the Ethernet interface of ISI's
Butterfly Gateway would lock up fairly quickly (in 40-60 seconds) at
packet rates of approximately 90 per second, forcing a gateway reset.
Transmissions therefore had to take less than this amount of time.
This problem has reportedly been fixed since the tests were
conducted.
In order to test the Wideband network under overload conditions, we
attempted several tests at rates of 5 1432-byte packets every 50
milliseconds. At this rate, the Wideband network ground to a halt as
four of the ten network BSATs immediately crashed and reset their
channel processor nodes. Apparently, the BSATs crash because the ESI
(Earth Station Interface), which sends data from the BSAT to the
satellite, stops its transmit clock to the BSAT if it runs out of
buffer space. The BIO interface connecting BSAT and ESI does not
tolerate this clock-stopping, and typically locks up, forcing the
channel processor node to reset. A more sophisticated interface,
allowing faster transmissions, is being installed in the near future.
Some more testing needs to be performed over the Wideband Network in
order to get a complete analysis of NETBLT's performance. Once the
Butterfly Gateway Ethernet interface lockup problem described earlier
has been fixed, we want to perform transmissions of 10 to 50 million
bytes to get accurate steady-state throughput results. We also want
to run several NETBLT processes in parallel, each tuned to take a
fraction of the Wideband Network's available bandwidth. Hopefully,
this will demonstrate whether or not burst synchronization across
different NETBLT processes will cause network congestion or failure.
Once the BIO BSAT-ESI interface is upgraded, we will want to try for
higher throughputs, as well as greater hardware stability under
overload conditions.
As far as future directions of research into NETBLT, one important
area needs to be explored. A series of algorithms need to be
developed to allow dynamic selection and control of NETBLT's
transmission parameters (burst size, burst interval, and number of
outstanding buffers). Ideally, this dynamic control will not require
any information from outside sources such as gateways; instead,
NETBLT processes will use end-to-end information in order to make
transmission rate decisions in the face of noisy channels and network
congestion. Some research on dynamic rate control is taking place
now, but much more work needs done before the results can be
integrated into NETBLT.
M. Lambert [Page 11]
RFC 1030 Testing the NETBLT Protocol November 1987
Although the raw bandwidth of the Wideband Network is 3 megabits per
second, currently only about 1 megabit per second of it is available
to transmit data. The large amount of overhead is due to the channel
control strategy (which uses a fixed-width control subframe based on
the maximum number of stations sharing the channel) and the low-
performance BIO interface between BBN's BSAT (Butterfly Satellite
Interface) and Linkabit's ESI (Earth Station Interface). Higher-
performance BSMI interfaces are soon to be installed in all Wideband
sites, which should improve the amount of available bandwidth.
Bandwidth on the Wideband network is divided up into frames, each of
which has multiple subframes. A frame is 32768 channel symbols, at 2
bits per symbol. One frame is available for transmission every 21.22
milliseconds, giving a raw bandwidth of 65536 bits / 21.22 ms, or
3.081 megabits per second.
Each frame contains two subframes, a control subframe and a data
subframe. The control subframe is subdivided into ten slots, one per
earth station. Control information takes up 200 symbols per station.
Because the communications interface between BSAT and ESI only runs
at 2 megabits per second, there must be a padding interval of 1263
symbols between each slot of information, bringing the total control
subframe size up to 1463 symbols x 10 stations, or 14630 symbols.
The data subframe then has 18138 symbols available. The maximum
datagram size is currently expressed as a 14-bit quantity, further
dropping the maximum amount of data in a frame to 16384 symbols.
After header information is taken into account, this value drops to
16,036 symbols. At 2 bits per symbol, using a 3/4 coding rate, the
actual amount of usable data in a frame is 24,054 bits, or
approximately 3006 bytes. Thus the theoretical usable bandwidth is
24,054 bits every 21.22 milliseconds, or 1.13 megabits per second.
Since the NETBLT implementations are running on Ethernet LANs
gatewayed to the Wideband network, the 3006 bytes per channel frame
of usable bandwidth translates to two maximum-sized (1500 bytes)
Ethernet packets per channel frame, or 1.045 megabits per second.
M. Lambert [Page 12]
RFC 1030 Testing the NETBLT Protocol November 1987