|
Introduction: Redefining the S/390 mainframe
 System structure
requirements for the S/390 platform transition
In 1993, IBM began revitalizing the S/390* platform in several
steps, with most of the transformation complete by mid-1997; key
elements of this process included the following:
- Reducing the cost of the processor by switching to CMOS chip
technology This made possible price/performance improvements of
35-40% per year. Since the performance capability of CMOS would not
accommodate customer workloads currently running on bipolar
(ES/9000*) systems, the capacity of the symmetric multiprocessor (SMP)
system and the engine size of CMOS systems were increased each year to
provide replacements for successive generations of bipolar systems
(Figure 1).
- Introduction of parallel systems To meet customer
requirements for higher availability, incremental increases in
capacity, and large-capacity scaling, IBM introduced a unique
clustering approach for S/390 (the Parallel Sysplex*)
[1,2]. This
approach confined the need to adapt to the cluster technology to the
operating system and the database and transaction monitoring systems.
This technology is now supported in all key IBM-supplied operating
systems and compilers and many independent software vendor products.
- Support of open interfaces This has included support for
connection interfaces such as Ethernet, FDDI, and ATM; offering TCP/IP
as an alternative network protocol to SNA; and supporting UNIX**
standard programming interfaces.
 Figure 1
Effectiveness of bipolar replacement
Performance
In the transition of S/390 systems from bipolar to CMOS
technology, rapid improvements in the circuit density and performance
of CMOS have allowed IBM to produce four generations of S/390 CMOS
systems in the past four years (1994-1997). The uniprocessor
"size," or effective throughput, of the last three generations
(9672-R*2, 9672-R*3, 9672-R*4, and 9672-R*5) delivers uniprocessor
speeds roughly comparable to those of the bipolar versions of the
3090-180J, 9021-520, and 9021-711 uniprocessors, respectively, as shown
in Table 1.
Table 1
Equivalence of CMOS and bipolar uniprocessors.
| Date | Generation
| CMOS model |
Equivalent bipolar model |
|---|
| 4/94 | G1 |
9672 R11 | 3090-180 |
| 7/95 | G2 |
9672 R12 | 3090-180J |
| 9/96 | G3 |
9672 R14 | 9021-520 |
| 6/97 | G4 |
9672 R15 | 9021-711 |
Reliability
In addition to the rapid improvement in processing capacity of the
S/390 CMOS-based systems, an even more dramatic improvement has been
made in the reliability of the systems. Reliability enhancement through
a) improved intrinsic failure rates of CMOS technology versus bipolar;
b) extensive reductions in the total number of parts required for the
CMOS systems (Figure 2); and
c) continued improvements in fault-tolerant design [redundancy, ECC,
cache-line delete, sparing, and N + 1 power (an
extra power supply)] resulted in an improvement of nearly two orders
of magnitude in MTBF (mean time between failures) of the 9672-RX5 over
a 3090-600J bipolar system.
 Figure 2
Scalability of performance
Regardless of processor technology, the scalability of a
multiple-processor system representing a single-system image (SSI) is
limited to a small number of processors. This is due to the well-known
fact that as the size of the single system increases, the cost of the
hardware and software components of the system increases prohibitively
in order to sustain scalability for a large number of processors. The
IBM Parallel Sysplex constitutes multiple computing images operating
under OS/390* in a system complex (sysplex). The sysplex uses the
coupling facility as a means of sharing data, resulting in a
high-performance parallel OS/390 operating system configuration. With
the advent of this IBM parallel sysplex technology, several processor
complexes, each in its own right a multiple processor, could be made to
appear as a single system so far as the end user is concerned.
With this transition to parallel sysplex technology, it is also
possible to maintain a near-linear degree of scalability even with
the use of several tens of processors.
Figure 3 shows this desirable scalability
characteristic for the parallel sysplex in comparison to the
scalability of a single multiple-processor system. In this study, the
LSPR (Large Systems Performance Reference) mixed
workload¹ is run on the single
multiprocessor system, pertaining to either bipolar or CMOS technology.
 Figure 3
The parallel sysplex comprising CMOS engines is used to run either
CICS/DBCTL or IMS/DB2 data-sharing EBM benchmarks. In the CICS/DBCTL
workload, the IBM CICS* subsystem acts as the transaction manager, and
the database manager is the IMS DBCTL. This workload contains
transactions from CICS applications. In the IMS/DB2 workload, the
transaction manager is IMS* at the front end, with DB2* V4 acting as
the database manager. The degree of data sharing in both of the
workloads is maintained at 100%. The scalability trend is
approximately the same for the two workloads, and continues to hold
when the parallel sysplex of CMOS G1 is replaced with G2, G3, or G4.
There is an initial cost for customers of about 10% in throughput when
two systems are enabled for data sharing in comparison to a
non-data-sharing single system. As more systems join the parallel
sysplex, each additional system incurs less than 0.5% cost in
throughput, thereby providing a high degree of scalability. Thus,
S/390 parallel sysplex scalability supports a very large commercial
data processing environment.
Continuous availability design point
In addition to the high-availability characteristics described
above which protect customers from unscheduled outages, many other
system design features improve the total system availability for
scheduled hardware maintenance (concurrent repair, such as N
+ 1 power supply channel cards, concurrent microcode patch apply,
etc.) Along with the exploitation of parallel sysplex technology, S/390
provides the most nearly continuously available computing environment
in the industry today. This is true especially from both hardware
and software standpoints, especially the latter. In a parallel sysplex,
not only can systems be switched in and out of the complex for hardware
problems or upgrades without disruption; the same is true for software
problems and maintenance. The customer can have near-continuous
availability by using the IBM S/390 CMOS servers and parallel sysplex
technology.
Evolution of mainframe servers and network systems
Resurgence of
mainframe servers
IBM mainframes were physically large in comparison with pre-1990
systems, expensive to buy and maintain, difficult to operate, and
limited to a host-centric domain. Without alternatives, large-system
computer users were bound to that environment. However, despite these
attributes, mainframes were still the marketplace's premiere platform
for mission-critical applications. As long as such applications were
profitable and provided a business advantage, IBM mainframes held a
positive position in the marketplace.
By 1993, however, alternatives to the mainframe were clear, and the
marketplace was ready to move. Client/server technologies were
appearing that were smaller, less expensive, and easier to maintain and
operate, with lower-cost, more productive, user-friendly, "total
solutions" application software. An impending mainframe extinction
was predicted. Reacting to this changing environment, IBM developed a
mainframe transformation strategy which is currently successful. As a
result, the resurgence of the mainframe is proceeding, although in a
form that differs dramatically from the mainframe of old. Today's
mainframe is smaller (Figure 2), less expensive
to buy, less expensive to maintain (Figure 4)
[3], and definitively more
open. The reduction in physical size is a result of our
transformation from bipolar to CMOS hardware and the associated change
from water-cooled to air-cooled technology. The increased density of
CMOS allows many more circuits per chip at a much lower power level,
which also requires less cooling. The creation of OS/390 has reduced
the cost of software and provided for openness that facilitates more
competitive, leading-edge "total solution" applications
[4].
Also, support provided for industry-standard connection devices such as
Ethernet, ATM, and FDDI has enhanced the attractiveness of the
mainframe. The parallel sysplex approach is the other key attribute of
this successful transformation, since parallel sysplex allows for the
clustering of far more processors in an integrated system than was
practical in previous SSI multiprocessor systems.
 Figure 4
Thus, the water-cooled bipolar mainframe of yesterday has been
transformed to an air-cooled CMOS large-system server that is well
positioned to meet the changing needs of the marketplace.
Requirement for
microprocessor growth with industry
The effect of combining competitive price/performance and open
interfaces has been to attract many new applications to S/390 that were
not originally written for it. Since many such applications are
ported from UNIX/NT platforms and are generally not optimized for
parallel sysplex, it is a key requirement to be able to scale their
performance. This requires the SMP including the microprocessors to
increase 40-50% [5] in performance
every year.
SMP design
targets--staggered development pipeline
Since the design of a new microprocessor in the industry takes
about four years, it is necessary to structure the system
differently from conventional designs in order to deliver 40-50%
growth each year. The approach IBM has taken is to employ two
microprocessor design teams that produce alternate microprocessor
"cores" every two years, each with a technology upgrade in the
interim years (Figure 5). The
I/O and memory subsystems for the system are refreshed every two years,
but staggered relative to the introduction of new microprocessor cores;
this avoids the introduction of a new microprocessor with a new SMP
environment, thereby simplifying the verification and architectural
correctness of the resulting system.
 Figure 5
I/O evolution
To support the increasing processor performance of the S/390 CMOS
servers and to provide more open, industry-standard interfaces, the
S/390 I/O subsystem had to increase the number of parallel, ESCON*, and
inter-system I/O channels attached to the server, and also had
to develop a strategy to provide direct attachment of the new I/O
interfaces. Increasing the total I/O bandwidth and connectivity to
channels and new I/O adapters required the introduction of a new
internal link called the self-timed interface (STI). The STI was
introduced on G3 servers and provides high bandwidth at distances of
the order of several meters. Since the STI requires relatively few chip
I/O pins, many STIs are provided, further increasing the bandwidth and
improving the connectivity. While the STI provided the required
connectivity and bandwidth, the attachment of more channels also
required their redesign and remapping into state-of-the-art chip
technologies. The newer chip technologies reduced the cost and physical
size of the channels while improving their reliability and
performance. Traditional S/390 I/O used external, channel-attached
devices to provide open, industry-standard interfaces. Now, S/390
servers (Multiprise* 2000) have integrated adapters to provide
Ethernet, token-ring, FDDI, and ATM interfaces. The ATM adapter is the
first to be attached using internal peripheral component interconnect
(PCI), and it demonstrates how newadapters may be attached in the
future. Disk storage was also traditionally attached using external
channel-attached devices, and the S/390 servers have developed adapters
to integrate Small Computer System Interface (SCSI) attached disk
drives. All of the changes described above have been made while
retaining the S/390 I/O programming model. Future directions may
augment this programming model to provide more direct I/O programming
interfaces.
Summary
The transformation of the IBM S/390 line of mainframe computers
has successfully reached a milestone: equivalence of CMOS-based and
bipolar-based microprocessors. The current success of G3 and G4 is also
due to the redefinition of the mainframe. The inclusion of key customer
requirements such as scalability and support of open interfaces has
made the new G3 and G4 offerings more competitive in the marketplace.
Parallel Sysplex and OS/390 offerings have set the stage for the
continued leadership and growth of S/390 products. This issue of the
IBM Journal of Research and Development provides insight
into some of the hardware development efforts that were an integral
part of the transformation.
*Trademark or registered trademark of International
Business Machines Corporation.
**Trademark or registered trademark of X/Open Co., Ltd.
¹The LSPR mixed workload is composed of equal
proportions of commercial batch, TSO, CICS, and IMS workloads.
References
Received May 20, 1997; accepted for publication July
24, 1997
|
HTML
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5