||Parallel Simulation Technology
Central JR Shinkhansen Simulator
Takashi Kawakami, manager of information system development at
Central Japan Railway Company, was pondering in 1992. It was time
to plan the next generation of their COMTRAC (COMputer aided TRAffic
Control) system, the Shinkansen (Bullet Train) control and monitoring
The Tokaido and Sanyou Shinkansen Lines between Tokyo and Hakata,
1,100 km (700 mi) long, are the busiest among all the Shinkansen
lines. On averrage, the lines carry 300,000 passengers everyday,
with a peak of 1 million per day, on about 1,000 trains at a speed
of 270 km/h (170 mph). This is something like 4,000 Boeing 747's
flying back and forth between San Francisco and Los Angeles. They
must increase the train frequency, keep the minuscule operation
error of 15 to 30 seconds, while maintaining their safety legend--no
passenger accidents in the thirty year history.
Central JR was wondering to what extent they could improve the
next generation system. They were also wondering how they would
test the new system and train the dispatchers. On the other hand,
they had had enough of long delivery and high cost of computer
systems and application software. These information systems, including
minor modifications, had always been contracted to a few mainframe
computer manufacturers. These systems had become black boxes for
the user, Central JR.
A new agent-based and complexity-based parallel computer and software
concept was introduced. Kawakami of Central JR quickly recognized
that his staff, without programming experience, could build information
systems on their own using the PIM and Paracell. They could test
what-if simulations on the do-it-yourself systems and evolve the
systems by themselves. These systems simulate operations as realistically
as the real complex world would be, giving proof to the plans,
developments, designs and schedules of the railway system.
Railroad system simulators are a natural application for the products.
With any conventional programming scheme and computer it would
have taken as much time and cost to build a simulator as a real
control system. Paracell, the programming language, and the PIM,
the parallel execution machine, made it possible to build such
a simulator at less than one third of the cost originally estimated.
Central JR has proved all of three major advantages of the PIM/Paracell
system with building the simulator. First, the simulation is real
time and the central control computer cannot identify which is
the real train operation or the simulator. Yaskawa Electric Corporation,
the supplier, did no application program coding at all. Application
programming was accomplished by domain experts - not programmers
- from Central JR. Third, the simulator easily scaled up from
the first small scale trials, ending up as a large scale system
to cover the whole distance of 1,100 km (700 mi).
Satisfied with the success of the simulator for their development,
Central JR is going forward to include the PIM/Paracell simulator
on on a much larger scale for the next generation Shinkansen Operation
Control and Monitoring (COMTRAC) system.
The Tokaido and Sanyo Shinkansen (Bullet Train) System
The Shinkansen operation started in 1964, over three decades ago.
Operations began running one Hikari super express and one Kodama
super express per hour. The Hikari and Kodama lines had the same
performance of 210 km/h (130 mph), but Hikari stopped at only
major stations and Kodama stopped at every station between Tokyo
and Osaka, covering 500 km (300 mi). Today, there are three classes
of bullet trains: Nozomi, Hikari and Kodama are running along
the Tokaido and Sanyo lines between Tokyo and Hakata, a distance
of 1,100 km (700 mi).
Nozomi has the fastest operation, running at a speed of 270 km/h
(170 mph). During peak hours, trains arrive and depart the Tokyo
station every three and a half minutes. The train frequency today
is one Nozomi, seven Hikaris and three Kodamas per hour compared
with just one Hikari and one Kodama 32 years ago. The JR company
is planning to add another Nozomi soon and then one more per hour.
Despite this heavy traffic, the system operates right on schedule
with a minuscule error of 15 to 30 seconds, most of the time.
There has not been a single passenger accident in the 32 years
history of Shinkansen.
Developing schedules involves months of work for quite a few railway
experts. They know what affects the train operations: available
trains, maintenance cycle, available drivers, different train
performance, stations, distance, signals, track switches, curves,
slopes, tunnels, bridges, and so on. Even extra trains must be
scheduled in advance, yet there is no way to run ad hoc trains.
Can they apply a big change all at once? No, when they started
the Nozomi 270 km/h super express, only four of those per day
could run only early morning and late at night. It took several
months before hourly operation of the Nozomi began.
There is a bottleneck near the Tokyo terminal station, the busiest
station. Since the Tokyo station has no hind tracks, arriving
trains and departing trains must cross over the main tracks. To
make things worse, the train yard is located before the terminal
station, not in hinterland, so that deadhead trains must travel
for a while on the busy main tracks. To reduce this problem and
increase the train frequency, Central JR is planning to build
another terminal station, Shinagawa, a few miles before the Tokyo
station. They must check the effects of the new station before
actual construction occurs.
On January 17, 1995, a big earthquake destroyed the Kobe Shinkansen
station and nearby tracks a few minutes before the train operation
of the day began. If it were after, a couple of trains with more
than 1,000 passengers could have crashed down to the ground. Japan
Railway barely escaped the death of their safety legend. The train
operation in this area had to be shut down for 100 days. JR was
losing US$15 million every day during that time.
The big earthquake in Kobe convinced Japan Railway to build the
second control center in Osaka in addition to today's sole control
center in Tokyo. The two centers will work as a back-up system.
When one center is on-line, the other is used to train the operation
center dispatchers. They need a simulator system to work with
the back-up system for dispatcher training.
The 300 km/h (190 mph) trains are already in test and will soon
be put into operation. They are planning even faster trains (200+
mph). New trains with different performance will make the already
complex system and its scheduling more complicated. There might
be undesirable emergent, chaotic, behavior. They need a simulator
system to check this.
If, for example, snow causes a delay, what happens? The whole
system goes out of control. Of course, there is no real danger
to passengers. The problem is that nobody knows the new schedules
even after the cause of the delay is removed. The train drivers
know only the original schedules instead of the new feasible ones.
Thus, every driver tries to catch up with the original schedules.
Soon, several trains are stuck following a train that stops at
a station. As soon as the leading train leaves the station, all
of the following trains start at once causing a peak power demand
to the power substation. They want to know when the flock of trains
hits the busiest Tokyo station. They want to model revised, seemingly
feasible schedules on a simulator system that performs in faster-than-real-time
Central JR sought to improve these situations to increase the
train frequency and to make the schedules more flexible. In addition,
they had several ideas for improved service. However, prior to
designing and implementing an eintirely new system, they have
to test the ideas. But, how? They required a simulator at a reasonable
cost that provided quick turnaround time, and the ability to experiment
with different scenarios easily. Conventional simulators are either
too simple to predict system behavior or too expensive and slow
to be practical.
Building a Simulator for the Shinkansen System
The Shinkansen tracks are segmented into1.6 km (1 mi) lengths
between stations. Lengths in stations and their neighborhood are
much shorter. The total number of track segments far exceeds 1,500
since the eastbound and westbound tracks alone are about 1,100
km (700 mi) each. Each track segment is the basic unit of control
for the train traffic and includes several relay logic elements.
Track segments are locally interlocked and generate signals for
the following segments. Some of the relay logic signals are sent
to the central control system every three seconds.
The central control system sends traffic control signals including
speed limit information and route control commands, i.e., track
switching command, based on the train schedules and tracking data.
At peak hours there may be 130 or more trains on the tracks, and
the shortest period between trains is three and a half minutes.
The number for the day is about 1,000 trains.
The Shinkansen simulator developed includes about 1,000 train
agents, 130 or more of which may be on the track at a time, and
2,000 track segment agents. These agents work in parallel. Simulating
this system with a conventional programming scheme and computer
would require a very large effort and a long time in systems design,
flowcharting, coding and debugging; the resulting simulation would
be brittle. In addition, it would be hard to divide the task among
many people and nearly impossible to complete it in a short period
The programming and execution scheme of the PIM/Paracell system
simulates teamwork with a bulletin board and a clock. The team
members are to get data only from the bulletin board and to post
results on the same board. They don't need to have time-consuming
meetings. The clock frees them from directly acknowledging their
communication with each other. Adding members does not increase
overhead. Each can concentrate on his own task, yet a concerted
job can be achieved through the bulletin board and the clock.
This programming and execution scheme of the Paracell/PIM system
mapped perfectly to the desired implementation. Programming individual
small computing cells, or agents, that are to run in parallel
and in synchronization allows for train and track agents to be
quickly implemented and scaled up as needed. The concept of communicating
between agents only through the common global memory supports
the need to have a single coherent image of the system status
for running the simulation. This model fits the structure of the
Shinkansen system very well.
The actual simulator system consists of nine Macintosh computers,
each housing 1,000 PIM cells. Communication between systems is
via Ethernet. One of the units works as the central machine and
eight others simulate the 39 stations and yards that make up the
entire Shinkansen system. The application program was coded by
JR engineers using Paracell, a rule-base near-natural declarative
language. They did not necessarily have programming experience.
Rather, they had knowledge about the problem they were trying
to solve, and ideas about how to solve them. The simulator is
their product, not the supplier's, and they have a strong motivation
to improve the product further.
Simulator for the Next Generation Shinkansen System
Central JR has proved all three major advantages of the PIM/Paracell
system by building the simulator. First, the simulation is real
time and the central control computer cannot identify which is
the real train operation or the simulator. When they want prediction
of train operations, the simulator can run even faster-than-real-time.
Second, application programming was accomplished by Central JR
people without programming experience. The number of lines of
code in a conventional programming language would likely have
been ten times more than the actual code in Paracell. Third, the
simulator ended up as a large scale system to cover the whole
distance of 1,100 km (700 mi). It was a simple linear effort scaling
up from the first small rial for a few miles near the Tokyo station
to a full-scale operational simulation.
Satisfied with the success of the simulator on the Mac-PIM, Japan
Railway decided to include a simulator subsystem in their next
generation Tokaido and Sanyou Shinkansen Operation Control and
Monitoring System. It will be up and running by 1998. The simulator
subsystem will be used for testing fault-tolerant central control
computers. In addition, it will be used to train the dispatchers
for both normal and emergency operations. Finally, it will test
be used to test many what-ifs conditions as a means of integrating
new technologies, equipment, and concepts into the Shinkhansen
operations. For this next phase, a PIM system based on the PowerPC
and VME technology will provide over 20,000 PIM cells for better
reliability and more than double the capacity fo the original
PIM-based COMTRAC system. With this capacity, they will be able
to expand the scope of simulation to inlcude such areas as power
supply and demand between trains and substations.
Railroads are complex. Yet the PIM/Paracell system has proven
capable of simulating the Shinkansen. This technology can describe,
simulate, and find solutions to many complex problems.