Goeller on Telecom
Traffic
Tackling The Non-Blocking Issue
Business
Communications Review, Feb 1989
What Do We Mean By Non-Blocking?
Everybody says they want a non-blocking PBX,
but nobody wants to use the same definition of non-blocking. Here
are four "definitions" I have heard recently:
-
A call can always go through.
-
You can always get dialtone, even if
everybody picks up the phone at the same time.
-
There is no contention.
-
The ability to provide every telephone in
a group with an immediate connection.
Let's compare the first of these with the
"official" definition based on the one in the IEEE Standard
Dictionary of Electrical & Electronic Terms. There, a non-blocking
network (matrix) is one in which any IDLE outlet can always be
reached from any given inlet under ALL traffic conditions.
Obviously, non-blocking doesn't count if the port you want to reach
is busy.
Call waiting and executive override are, not
mentioned in the IEEE definition. They provide ways to let you
reach your party, even if busy, and work quite well even though,
under the circumstances, a non-blocking matrix would be no help at
all.
The second “definition” is apparently
confused with "mean time to dial tone," a completely different
traffic measure that used to be quite important to manual
switchboards and Strowger Step by Step switches, but has little
relevance in terms of modern COs or PBXs and has nothing to do with
whether or not the matrix is non-blocking.
But the concern is real. Suppose you are
operating an outbound telemarketing application and you specify you
want NO dial-tone delay to your agents. Your vendor could adjust
the system software to let all agents listen on the time slot that
returns dial tone when they come off hook. Everyone could get dial
tone very quickly, but the system just might not have enough DTMF
receivers to go around. For the specification to be meaningful, the
system would have to have one DTMF receiver for every line or trunk
that might originate a call, a solution that would be technically
absurd and ridiculously expensive.
Perhaps the relationship between dial tone
and the non-blocking matrix comes from electronic hybrid key systems
(actually small electronic PBXs) used behind Centrex systems. When
you pick up a phone in such systems, the dial tone you hear comes
from the Centrex switch; each Centrex line appears to the hybrid to
be a DTMF receiver, and if everybody picks up at the same time, a
probability which gets higher as the size of the hybrid decreases,
only a non-blocking matrix stands between you and the Centrex line.
However, the Centrex switch itself is not non-blocking; it plays the
averages in terms of its own DTMF receivers and just might, in a
busy hour, produce a modest dial tone delay.
The contention “definition” is also
unrealistic. Even on a probability basis, contention for a path
through a PBX switching matrix is seldom a factor compared with
contention for lines and trunks. Indeed, one of the best reasons
for using a PBX is to provide a means for allowing callers to
contend for expensive facilities. If having one DTMF receiver for
each line that might want to originate a call is absurd, consider
the need for having one trunk in each trunk group for every line.
This would avoid contention, but at what a cost! Traffic
engineering principles help configure PBXs so that a good grade of
service is provided without incurring the expense of facilities that
have a very low probability of being used.
The advocates of LANs make frequent reference
to contention in PBXs. But what they do not say is that LANs are
designed on a delay basis, buffering packets in a queue until the
shared channel is free. If your packet collides with another
already in transit, LANs institute additional delay with elaborate
retry procedures. As the LAN grows, these delays can become quite
noticeable. And on top of everything else, many LANs are designed
to block new packets from entering the system until old traffic is
cleared.
But this is not really the point. Whether
you are using a LAN or a PBX to provide 500 users with access to 30
ports on a mainframe computer, the computer can only handle 30
simultaneous dialogs. If a thirtyfirst user tries to connect to the
computer, neither LAN nor PBX, even if completely non-blocking, can
do anything to help. Once again, a non-blocking machine cannot
guarantee a connection when there is nothing available to connect
to.
Our fourth “definition” comes straight out of
a Rolm system description — the ability to
provide every telephone in a group with an immediate connection. If
we are looking for this ability, the only way we can be satisfied is
when some of the connections are ones the callers do not want, such
as connections to station busy or all trunks busy tone. Further,
referring to any connection as "immediate" is optimistic because
there will always be some delay until the processor gets around to
everyone, even if that delay is measured in milliseconds.
Rolm is not alone in creating “definitions.”
Northern Telecom is now advertising that its Meridian SL-1 is
non-blocking. Since the SL-1 is NOT non-blocking, it pays to
understand what Northern is actually talking about.
The SL-1 has individual line groups which can
serve up to 160 ports. Each line group is connected to the central
switch or group selector by a multiplexed link which has 30 time
slots. This means that no more than 30 ports in any given line
group can be in use at any one time.
At the group selector, a talk channel from
one line group can be connected to the listen channel to the same or
a different line group in each time slot. The total number of (one
way) connections possible is the number of multiplexed links
supported times the number of time slots per multiplexed link. But,
in early SL-1s, this number could not always be realized. The SAME
time slot had to be used for both the calling and the called party,
so that the group selector would make the right connection. If time
slot 24 were the only one free in my line group, and time slots 17,
22 and 28 were the only ones free in yours, we could not be
connected together for a conversation. We would have what Northern
calls a "failure to match," and the group selector would block the
call.
In all the current Meridian SL-1s, Northern
Telecom has added a time slot interchanger (TSI) in the talk side of
each multiplexed link. It changes the time slot assigned to me to
the one assigned to you and then, in your time slot, makes the
connection from my talk-side multiplexed link to your listen side.
This eliminates the failure to match, makes the group selector in
the SL-1 non-blocking, and greatly improves the SL-1's traffic
handling capacity. If you have no more than 30 ports in any line
group, the SL-1 is non-blocking.
Another Misused Term
Full availability (sometimes called full
access) is a related term frequently misused as a synonym for
non-blocking. Some brochures even claim that a PBX switching matrix
with full availability can handle 36 CCS per port. The IEEE
definition of availability is a little obscure, but an example will
clarify the point. In the days of Step by Step switches, with ten
outlets per level, it was frequently necessary to access trunk
groups with more than ten trunks. One way to do this involved
"grading," which meant that the switches were divided in to several
groups, and each group would have access to only ten trunks in the
larger group. That is, it had an "availability" of 10. For 50
trunks, what one might do would be to arrange the switches in 9
groups, with the first five terminals on each group having their own
trunks (total: 45), and the last five terminals sharing the
remaining five as an overflow group. A call would thus have a shot
at ten trunks and, with a uniform distribution of traffic, the
trunks would be well loaded. But there would always be the chance
that some particular call would find all its trunks busy while some
other switch group would have some of its trunks idle.
The more sophisticated hunting algorithms,
available first in crossbar systems and then in computer controlled
PBXs and CO switches, eliminated the need for grading and made it
possible for any trunk group to be fully available to any call.
This greatly improved the utilization of the trunks but,
unfortunately, there was still the chance that the switching matrix
itself might not have a free path to the desired output. In short,
a system could have full availability, but still not be
non-blocking.
This makes the distinction between full
availability and non-blocking clear: in a full availability switch,
any inlet can reach any free outlet in the absence of other traffic,
while in a non-blocking switch, any line can reach any free outlet
no matter how much traffic the system is already carrying.
Even a truly non-blocking switch can
experience a lack of full availability. For example, a PBX can be
configured with ten outgoing, ten incoming and ten two-way trunks.
The two one-way trunk groups hunt from opposite directions into the
two-way trunk group. This guarantees some outgoing channels when
incoming traffic is heavy (and vice versa), but it also means that
calls attempted in one direction may encounter an all trunks busy
condition when trunks in the other one-way group are idle.
The Costs of Non-blociking
Although we can easily design a non-blocking
matrix, there is no practical way to build a non-blocking SYSTEM:
one DTMF receiver and one trunk in each trunk group for each line,
and no called party ever busy. Thus system designers have, in the
past, reasoned as follows:
In a PBX, we assume six CCS per line (that
is, each line is assumed to be busy for 600 seconds, or ten minutes,
at random, in the busy hour). Put another way, the probability
that a given line is off hook is 1 in 6, or 0.167. Trunks have a
higher occupancy than lines (about 30 CCS, 3000 seconds or 50
minutes out of an hour), but because any trunk in a group will do,
we arrange to have a probability of something like 0.05 that at
least one trunk will be free to accept a new call at any given
time.
Given these probabilities, how much should be
spent to guarantee that the probability of matrix blocking is 0.01
or 0.001 or 0.00001? Or truly non-blocking? Or, to put it another
way, if a non-blocking matrix cannot guarantee a connection, why
should we bother?
Clearly, if non-blocking costs very much (as
it did in the days of electromechanical switching), obtaining it is
not worth the effort, in view of the probability that the terminal
we want to connect to is already in use. Indeed, a current book by
a Bell Labs engineer suggests that designing a non-blocking local
central office might be illegal, in that it would require customers
to pay for something they could never use. (So much for Centrex
being non-blocking.)
Who Needs Non-blocking?
There are, however, situations when
non-blocking switching capability is needed--ACD operation and
tandem tie-trunk switching being two of the most obvious. In both
instances, we tend to have large groups of servers, and each server
presents its matrix port with very high occupancy. But often, such
switches also have a full complement of regular PBX stations with
much smaller probability of being in use. When this is the case, we
quickly see the advantages of having a non-blocking matrix: we can
eliminate traffic balancing.
Most large PBXs, including those shown in
Exhibit 1, are composed of several line groups, usually
concentrating traffic to a small number of high traffic links to a
group selector. Their design is based on the probability that not
all ports will be off-hook at the same time, and larger size and/or
greater economies are possible by taking advantage of
concentration.
The group selector in these switches
establishes connections from one line group to another, or sometimes
between callers on the same line group. Because the links from the
line groups to the group selector are heavily used, the group
selector itself must approach or be a non-blocking switch.
However, when the line groups concentrate
traffic to the group selector, there is always the possibility that
more ports on one particular line group will be involved in
originating or terminating calls than there are time slots
available. If we put all the heavy users (ACD agents, tie trunks,
CO trunks, whatever) on one line group, and all the light users
(regular PBX stations) on the others, we are asking for trouble.
The line group with the heavy users will experience a high level of
blocking.
We have to balance the traffic by moving some
of the heavy users to other ports in light traffic line groups,
replacing them on the heavily used line group with people who use
their phones less. By evening out the traffic on the links to the
group selector, we increase the probability of finding a path
available out of the formerly overloaded group. If this is not
sufficient, we can add line groups and spread the traffic more
thinly.
On the other hand, if the switch had been
designed to be non-blocking, typical of the PBXs in
Exhibit 2, any
line, trunk, ACD agent, etc., could have been put on any matrix port
without regard to traffic considerations. Eliminating traffic
balancing and the administration necessary to handle it over the
life of the switch can, in many instances, turn out to be a real
bargain.
The InteCom IBX was one of the first switches
designed in this way. Each line group is connected to the group
selector via broadband talk and listen links on either optical fiber
or coax. Each port on a line group has its own permanently assigned
time slot. Because of the huge bandwidth available, there is no
coast advantage in minimizing the number of time slots.
The group selector can connect any talk path
to any listen path, making different connections in each time slot.
Group selectors of this sort are typically Time Slot Interchangers
composed of conventional RAM memory, usually two 8 bit bytes per
port. The maximum RAM needed for switching (and various related
support functions) is typically less than 256 Kbytes for a 10,000
port system. At today's prices, this is hardly a major cost item,
and tips the tradeoff between traffic balancing and non-blocking
heavily toward the latter.
Who Doesn’t Need It?
As exhibit 2 shows, several smaller systems
such as the Siemens Saturn and the Harris 20-20 also take advantage
of this technology, typically using 32 channel multiplexed links
with a time slot per port to a non-blocking group selector.
However, the majority of the under-200 line systems, typified by
those shown in Exhibit 3, are not non-blocking, but do not require
traffic balancing, either. These systems can be visualized as a
simple crossbar switch: lines and trunks come in on the verticals,
and connections are made via the horizontals. The horizontals are
usually time slots today, but the technology is unimportant. As
long as there are half as many horizontals (paths through the
switch) as verticals (lines, trunks, other ports), the PBX is
non-blocking.
For instance, an analog Mitel SX-100 or 200
with 31 matrix paths is non-blocking to 63 ports (the odd port does
not matter, because it has nobody to connect to when everybody else
is busy). Similarly, a Siemens SD-192 is non-blocking to 97 ports.
Both of these actually use a solid-state matrix with space division
cross-points. With time division, The Dimension 400 is
non-blocking to 129 ports and the Dimension 600 to 257. Systems 25,
75 and the single module 85 are non-blocking to 231, 472 and 513
ports respectively. A Rolm single-node CBX with ROLMbus 74 can go
to slightly over 300, and an Alcatel System 3100 is non-blocking to
181 ports.
What happens if we increase the number of
ports? The probability that more conversations will be required
than we have horizontals will increase slowly. When we have added
enough ports so that the traffic generated makes the probability of
finding all horizontals busy greater than about 0.001 (one in a
thousand), we stop, or get a bigger system.
None of these systems needs traffic
balancing; they are arranged so that all lines and trunks have equal
access to the total group of horizontals. The horizontals are
accessed on a full availability basis, meaning that everybody has
the same chance at them. No horizontals are available for the
exclusive use of certain lines or trunks. Therefore, traffic
balancing is neither required nor possible in these systems, even
though at larger sizes they are not non-blocking.
Conclusions
To summarize, the primary value of a
non-blocking matrix is to eliminate the cost of traffic balancing
over the life of the PBX. Modern switching techniques have greatly
reduced the cost of making a switch non-blocking in the first place,
so we might just as well. At smaller sizes, however, architectures
are common that do not need traffic balancing even though they are
not non-blocking.
Now that you understand non-blocking,
availability, and traffic balancing, you can separate your business
acquaintances (whether designers, vendors or customers) into two
groups when they begin talking about non-blocking PBXs: those who
know what the words mean, and those who are glibly spouting
zip-terms in the hope of impressing the ignorant. Knowing what the
words mean will also help you make the right choices to meet the
needs of your company or your client. You can act with the
confidence of competence rather than with hope produced by mouthing
magic mantras.
[
Top ] [ Next Chapter ] [
Table of Contents ] |
