Over time (ref. graph ), many of of our customers will move to the new network centric computing paradigm. In this environment, the increasing functional complexity, computational burden, and business criticality of our server solutions means that customer expectations with respect to the raw performance and technical sophistication of our server solutions will rise. It also means that customers will come to expect near continuous availability and instant recovery from faults. What then are some of the base elements or components that IBM must provide to serve this vision? What will the dominant servers in this future market look like?
Technology and market forces are driving the server industry toward a server
model that can be expected to dominate the industry. The
characteristics of this future can be articulated as:
It will be a shared memory multiprocessor (SMP). The throughput of a single SMP node will continue to
grow very rapidly, driven by single processor performance that doubles every 18 months, and by an ability
of newer SMP's to handle an increasing
number of processors. The number of processors which can be effectively
utilized is now growing by approximately
a factor of two per processor generation. As a result single SMP node performance is now growing by
more than a factor of 2 every year.
Clusters of SMP nodes will of course play a very important role when capacity and
throughput demands exceed the capabilities of a single node, but the
vast segment of the
available market revenue will be addressed by the single SMP server form, and success
in this SMP market will be essential to success in the cluster
market.
It will have very large memories, requiring 64-bit processor technology.
Within a $1M server hardware budget there will be adequate resource
to buy 64 gigabytes of mainstore memory.
The transition to 64-bit technology will be the fastest transition of this type
yet seen. Even modest desktop servers will break the 4 gigabyte barrier
as memory prices drop below $5K per gigabyte.
Large durable mainstores will be used as a stable storage medium for frequently
accessed customer and systems data, becoming the highest level of the storage hierarchy.
The first (very successful) attempts to exploit semiconductor memory as a stable
storage medium have already been made in high end disk controllers and solid state disks.
These provide the nonvolatility and fault protection demanded by this new role.
SMP mainstores will adopt nonvolatility and fault protection technologies as the
cost advantage of moving semiconductor based stable storage closer to
the processors become advantageous.
Extensive use of fault tolerant technologies will be made within a single SMP.
As SMP's become larger, and as more customer data is stored in-memory,
impact of an SMP failure simply becomes unacceptable for routine failures.
With
increased scale, the frequency of routine failures grows and the
recovery time
to an alternate SMP also grows. Approaching several hours for large
systems, this time becomes too disruptive
to be acceptable. Remote site and disaster recovery will also become
increasingly
important, but will continue to remain too slow, too disruptive, and too
costly to be invoked except for extraordinary failure scenarios.
The focus industries where reliability and responsiveness have historically set the most demanding expectations are finance, telco, healthcare, retail, manufacturing, and warehousing. With the expected explosive growth in internet based electronic commerce, we can predict that the finance and telco areas will remain as the largest and most demanding industry sectors.
For more
information
send e-mail to Webserve.
Last modified 1996 July 12
