| Here is the document refered to in an earlier reply. I moved the Pyramid
section to the front but left the other material for reference.
ALPHASERVER 8400 (TURBOLASER)
FOR
DATA WAREHOUSE
COMPETITIVE ANALYSIS
Digital Equipment Corporation Internal Use Only
July 1, 1995
Introduction
The following document was prepared to help you compete against the many other
vendors in the high end data warehouse market arena. This was developed from
the high end view point, i.e. 100 GB and higher data warehouses. Systems with
both SMP and MPP architectures are represented.
The format is a description of each vendor's hardware and capability followed
by a brief strengths and weaknesses section. This is followed by a "selling
turbolaser against" discussion. This document enumerates the important
differentiators for participants in the high-end DBMS marketplace. The last
two pages are a reference summary of strengths and weaknesses.
This document is for Digital Internal Use Only!
7. Pyramid Technology Corporation
Pyramid provides both SMP and MPP
systems specifically targeted at the high-end marketplace. The Nile 150 SMP
can be configured with up to sixteen R4400 CPUs and 4 GB of primary memory.
Pyramid has recently upgraded to the 200 MHz R4400 with a SPECint92 rating
of approximately 135 per CPU. Pyramid has been providing SMP solutions
to the high-end marketplace for over ten years, and has some of the best
references in the industry. They were one of the first UNIX vendors to
implement Oracle Parallel Server in a clustered SMP environment (with mode 3
capability), and currently allow up to a four node cluster with each node
consisting of up to sixteen CPUs. Pyramid supports Oracle, Informix, Sybase,
Ingres and other DBMS products - but is clearly biased toward Oracle.
Over eighty percent of all Pyramid installs run Oracle. The weak point in the
Nile 150 SMP architecture is found in the extend bus architecture which
peaks at 100 MB/s in I/O bandwidth using a maximum of four
extend buses at 25 MB/s each (actual sustained I/O bandwidth is around 80
MB/s. The system bus is 400 MB/s and is rarely the bottleneck as compared to
the I/O extend buses in a high-end DBMS application. The DC/OSx operating
system (an AT\&T System V Release 4 implementation) is quite mature and
scales well up to at least 12 CPUs.
The RM1000 MPP product uses the same operating system and CPU chip as the SMP
offering to preserve complete compatibility between the products. The
RM1000 is configurable up to hundreds or thousands (theoretically) of
individual nodes, each of which is a single R4400 CPU with up to 256MB of
memory and the DC/OSx operating system. Each node has two F/W SCSI-II
controllers and nodes are connected via a high bandwidth, low latency grid
interconnect network. The RM1000 runs only Oracle to date, although other
ports are in various stages of planning or implementation. The SMP product
can be"plugged into" the MPP product as a node, and therefore Pyramid has a
strong story regarding investment protection for their customers who
can start out with low-risk SMP technology and migrate into MPP technology
when application requirements demand such a move. Pyramid has substantial
expertise in "glass house" environments and differentiates itself with
ESCON connectivity, full function backup and tape management systems, and
other production quality infrastructure. Pyramid was only a $250M company
up until its acquisition by SNI, and therefore is not well known within
the industry. Analyst reports from the likes of Gartner Group, Meta Group,
IDC, etc. tend to ignore Pyramid. The absorption into $30B SNI may or may
not change this situation.
Strengths: Oracle relationship and quality of Oracle port, completeness
of product line with both SMP and MPP solutions, seamless migration
pat from SMP to MPP, commercial references and experience, superior
benchmarking expertise, operating system scalability.
Weaknesses: Weak I/O subsystem, low visibility in the marketplace,
impact of SNI acquisition is unclear - but SNI has bad reputation in such
scenarios.
Selling TurboLaser Against Pyramid Technology Corporation:
Emphasize price-performance. Pyramid is almost always at the high-end
of the pricing chart (surpassed only by AT\&T and Tandem). They
have a good Oracle relationship, but do not have 64-bit capability
in their operating system. The DC/OSx version of UNIX that Pyramid
runs is SVR4, but has not been explicitly certified by very many
third party software vendors. Pyramid has a strong story in migrating
from SMP to MPP within their platform, but a TurboLaser implementation
would allow deferring MPP for much longer than a Nile (or Nile Cluster)
would allow because the TurboLaser is a much more powerful SMP than the
Nile 150 SMP. Also point out that Pyramid's future is somewhat unknown,
based on the recent SNI acquisition. SNI has a bad track record in
successfully integrating its acquisitions into the company. SNI has been
hands off Pyramid up until now - but who knows about the future?
1. AT&T Global Information Solutions
The ATT GIS SMP is an Intel based implementation with up to eight CPUs per
node. Sixteen CPUs per node has been announced. Up to four nodes
can be clustered in either the 3600 or 3550 architectures. The 3600
architecture also provides Teradata capability for DSS ``below the Y-Net''.
The Y-Net allows a cluster of UNIX SMPs (referred to as Application
Processing Nodes to handle open systems OLTP and other workloads
(usually Oracle DBMS), and uses a variety of trick-feed or batch update
techniques to populate the Teradata system for DSS in what they refer to
as the ``Information Factory''. The Teradata is a special-purpose database
that runs on a proprietary Intel hardware implementation using a
special-purpose operating system known as TOS (Teradata Operating System).
The Teradata is a shared-nothing hardware and software implementation
that has been around for over ten years and is well-proven as an MPP DSS
database with reference sites in the multiple terabytes range.
The Teradata was a monopoly in high-end database solutions for DSS
until just a few years ago when Oracle entered into the fray with its
Parallel Query implementation in Oracle 7.1. Since then, DB2/6000 Parallel
Edition, Informix, and Sybase have also come into the DSS marketplace.
The high margins that the Teradata used to command have rapidly been
shrinking in reaction to this increased competition. It will
be difficult for the Teradata to compete effectively with the merchant
database vendors because it does not have sufficient installed base to
allow amortization of development costs for new features and performance
enhancements. ATT has addressed the undesirability of the special-purpose
hardware implementation by porting the Teradata software into a UNIX
environment for release sometime this year. The port will allow the
Teradata to run on a general-purpose processing
architecture rather than as an SQL co-processor as it previously executed.
The port will not be available until release of the ``Common Node''
architecture which is the 1995 incarnation of the long awaited 3700 product
(which was originally scheduled for release in 1993). The Common Node
architecture will provide an MPP architecture with a scalable interconnect
(the Y-Net is scalable in connectivity, but not in bandwidth) and SMP nodes.
In the interim, prior to general availability of the Common Node, ATT has a
lot of issues in its product line. Neither the 3600 or 3550 SMPs scale well
past four to six CPUs, due to operating systems issues. The 3600 is not
able to provide Oracle with sufficient I/O bandwidth in a high-end clustered
implementation because there are not enough slots left in the backplane of
the machine to support an appropriate number of I/O controllers for large
disk configurations (because they are taken up by the Y-Net
interfaces). The 3550 does not support the Teradata, so it does not use the
Y-Net and therefore has better I/O bandwidth capability (more slots for I/O
controllers in a clustered environment), yet it does not upgrade easily into
a Common Node configuration.
Strengths: Teradata has proven (production) references and scalability.
Weaknesses: Poor scalability in UNIX SMP implementation, high pricing,
Intel chip set, insufficient I/O capability for Oracle in clustered 3600
architecture, no upgrade path for 3550 into Common Node architecture,
customers do not want proprietary Teradata database.
Selling TurboLaser Against AT&T Global Information Solutions:
Emphasize open systems commitment, price-performance, and raw CPU power.
In the data warehouse marketplace, AT&T GIS will almost always push
the Teradata DBMS. DEC should sell against this solution by emphasizing
the superior functions and features of the Oracle database versus Teradata.
The Teradata DBMS is years behind Oracle and is unlikely to ever catch up.
The fact that Teradata is a special purpose database means that it will
never have the installed base of Oracle - Teradata has hundreds of active
installs whereas Oracle has many tens of thousands of active installs. The
economics of this equation mean that AT&T GIS will NEVER be able to justify
investment in the Teradata DBMS to the same extent that Oracle can. Basic
features like outer join are only now in beta test in Teradata environments,
whereas this has been in both the ANSI SQL standard and in Oracle for years.
Stored procedures, advanced arithmetic functions, triggers, decodes, etc.
are all important features that are in Oracle today and are unlikely to
ever appear in Teradata. Teradata had a monopoly on parallel database
capability that was broken when Oracle introduced Parallel Query in
its 7.1 release of 1993. Since that time, Teradata has had to lower its
prices dramatically to compete and still has a tough time getting new
accounts (most business today comes from upgrades to existing installs).
Teradata has been around for a long time, and its antiquated DBMS
implementation shows it: the Teradata DBMS is still implemented using a 16-bit
software architecture whereas the new Oracle release on TurboLaser makes
use of a 64-bit software architecture! This has large implications in how
the DBMS is able to make use of memory resources for efficient implementation
of sort work area, database buffer caching, and software pipelining.
From a hardware perspective, DEC TurboLaser is miles ahead of AT&T GIS.
AT&T uses the underpowered Intel chip, as compared to the Alpha chip in
DEC TurboLaser. A single TurboLaser CPU is worth (conservatively) three or
four AT&T CPUs. This leads to enormous pricing benefits for TurboLaser in
equivalent machine configurations. In fact, the current Teradata systems
are still shipping with 486 chips - not even Pentiums! The AT&T GIS
Clustered SMP systems (above the Ynet) are Pentium based, but scalability
on a per SMP node basis is limited to four to six CPUs - less than a two
CPU TurboLaser configuration! Have the customer check AT&T GIS references
for the Clustered SMP to verify these scalability figures. AT&T GIS will
be increasing the number of CPUs that can be configured into an SMP node,
but it does little good for the customer because the hardware and software
implementation will not provide for scalability beyond the four to six figure
anyway. AT&T GIS is almost always priced at a premium, and the hardware
superiority of TurboLaser will widen the price-performance gap even further.
2. Cray Research SuperServers
The CRS 6400 is a 64-way SMP based on the
SPARC chip. The system will support up to 16 GB of memory and over 2 TB
of on-line disk with up to 64 I/O channels. The CRS 6400 architecture
involves a multiple bus system that yields a theoretical peak of 2.6 GB/s
of bus bandwidth; each of the four system bus structures (XDBUS Quartet)
provides 660 MB/s. Actual peak performance is 1.76 GB/s on the combined bus
structures. The CRS 6400 runs an ABI (Application Binary Interface)
compliant version of Solaris, so all applications that will run on a Sun
workstation will run on the machine (although they will not necessarily
parallelize). CRS has been around with a production port of Oracle for
only about a year. Recently, they have been giving Informix a lot of
attention in their DSS bids (but still use Oracle in most OTLP bids).
Each SPARCPlus processor delivers 84 SPECint92 in integer performance. CRS is
likely to move to the faster clock rate SPARCPlus chip in the near future.
Clustering capability is not yet available, but is likely soon. CRS is a
fully owned subsidiary of Cray Research, Inc. in Minneapolis and focuses
only on commercial computing (both OLTP and DSS). Recently, a top management
re-haul was done at CRS and many people were pulled in closer to the parent
company with a reorganization to provide tighter controls from Minneapolis
(likely result of poor profitability). A joint venture deal with Amdahl
allows Amdahl to OEM the CRS 6400 and provide all customer support for the
box.
Strengths: Aggressive pricing, lots of third party software, good scalability
up to at least 30 CPUs (particularly with OLTP).
Weaknesses: Minimum configuration is 16 CPUs, no production reference sites,
future of company in doubt, poor showing in DSS marketplace, poor
availability strategy (clustering not yet available), lack of support
structure (although teaming with Amdahl is helping to address this
issue).
Selling TurboLaser Against Cray Research Superservers:
Emphasize raw CPU power and company viability. The CRS 6400
will provide up to 64 SPARCPlus CPUs in an SMP configuration. This would
be the equivalent of a 16-way TurboLaser configuration (if such a
configuration existed). However, the reality is that the CRS 6400 scales
only to (approximately) thirty-two nodes due to database limitations - the
equivalent of an 8-way TurboLaser configuration. TurboLaser will scale
beyond 8 CPUs and will be more cost-effective in doing so in the high-end
range. Cray will be very aggressive with pricing, but TurboLaser should
ultimately win out in price-performance (but it will not be by a large
margin unless deep discounting is applied). To sell against Cray, emphasize
the importance of powerful CPUs (TurboLaser operates within scalability
limitations of the database because it needs fewer CPUs to do the same
job) and the depth of the DEC organization. Cray Research Superservers
is is a small organization operated within Cray Research, Inc. (of
supercomputer fame) and does not have the strong engineering and support
organization that DEC enjoys. It is also unclear if Cray can make
an effective transition from scientific computing to commercial computing.
CRS was formed to address the commercial marketplace, but its success has
not been overwhelming - profitability issues exist and a recent purge of
top executives within CRS occurred as the organization was pulled in
closer to its parent company whose historical focus and understanding
is targeted at scientific (not commercial) data processing. Staying
power of this start-up company should be questioned in the customer's
mind.
3. Data General
Not really a player in high-end SMP. Nobody has confidence that
DG will be in business long enough to provide good support
for the Avion series, even though it is a reasonable machine design. Biggest
impact of Data General is through Clarion RAID arrays.
Strengths: Good price-performance via aggressive pricing.
Weaknesses: Poor marketing, vendor viability.
Selling TurboLaser Against Data General:
Data General will be extremely aggressive in pricing, but TurboLaser
price-performance should yield parity for low-end configurations
and has the capability to scale up to much larger configurations with
excellent price-performance. Also emphasize the relationship with
Oracle and the first 64-bit DBMS implementation. Doubts about
vendor viability should also be emphasized. It is unlikely that DEC would
ever lose a deal to Data General.
4. Hewlett-Packard
The Hewlett-Packard T500 is an SMP with up to eight PA-RISC CPUs (12-way
capability has been announced). Peak bandwidth is 800 MB/s on the
system bus. The PA-RISC 7150 provides 136 SPECint92 in performance per CPU.
The PA-RISC 7200 provides 200 SPECint92 in performance per CPU. The HP-UX 9.x
operating system has scalability problems beyond four to six CPUs in a
high-performance DBMS environment because (among other reasons) all I/O
interrupts are farmed to CPU 0. This CPU soon becomes the bottleneck
and adding more processors yields only marginal performance gains. The HP-UX
10.0 operating system that will be out in June should fix this problem, but
scalability to twelve CPUs is in doubt. Hewlett-Packard aggressively pursues
the Data Warehouse opportunities with its ``HP OpenWarehouse'' product.
This product consists of some HP developed and integrated software which
includes: (1) PRISM Warehouse manager for data extraction and transformation,
(2) Red Brick, Informix, Oracle, Informix, Ingres, or Sybase database
engines, (3) Information Advantage AXSYS for custom development of DSS, EIS
and MIS applications, (4) HP Information Access which is a proprietary tool
for providing point-and-click access to the DSS database engines, and
(5) IBI's EDA/SQL for gateway access to legacy data sources.
Hewlett-Packard tends to favor Red Brick since this database engine runs well
on powerful CPU systems (DEC is the only vendor that can beat an HP CPU) and
does not exploit parallelism very well. Recently, Hewlett-Packard has
switched much of its merchant relational database attention from Oracle to
Informix (since Sequent no longer has exclusive rights to the Parallel Query
capability in Informix). Hewlett-Packard used to own 25\% of Informix, so
there is a lot of cross-pollination between the two organizations and they
tend to team with Informix quite often. A clustered implementation of the
T500 is available and will go up to 32 nodes (each node can be a 12-way SMP)
using an Oracle DBMS.
Strengths: Powerful CPU, very aggressive price-performance, superior
service reputation, strong open systems commitment, good
partnership with third party software providers.
Weaknesses: Poor scalability, RISC chip at end of lifetime, Intel
partnership is bad signal for high-end market.
Selling TurboLaser Against Hewlett-Packard:
Emphasize scalability of the TurboLaser in terms of processing power
delivered. Hewlett-Packard has no credible MPP strategy and its
SMP has scalability problems in the HP-UX implementation (no more than
four to six CPUs worth of scalability). HP-UX 10.0 available this summer
should address some of their scalability problems, but it is not possible
for an unclustered T500 to scale up as high as a TurboLaser in terms
of raw processing power. Hewlett-Packard will attempt to address this
deficiency in the data warehousing environment by selling the HP
OpenWarehouse solution using multiple T500 systems on
the back-end. Point out to the customer that the multiple T500
solution is actually a distributed database solution to the warehousing
requirement and joins across databases are very inefficient. Hewlett-Packard
will claim that you can design a database whereby joins are not required
across the T500 systems using date segmentation or other techniques. This
is absurd, and customers should be reminded of the importance of flexibility
in a data warehousing environment. Having a single TurboLaser box to address
data warehousing requirements is much more effective in terms of
flexibility, price-performance, and systems management. Similar arguments
hold true for OLTP workloads where Hewlett-Packard will be required to
propose multiple SMP boxes to do the same workload that a single
TurboLaser can achieve.
5. International Business Machines
IBM has a vast hole in its product line between the 4-way RS/6000 SMP and
the SP2. The 4-way SMP is woefully underpowered for high-end applications,
yet customers do not want to be forced into MPP prematurely. IBM routinely
gets slammed by Meta Group and other analysts for their tendency to pitch MPP
for very small systems. The IBM clustered solution with HA-CMP has just
recently gone mode 3 (makes active use of multiple nodes
rather than hot stand-by), but still allows only uniprocessor nodes. Most SP2
installs are Oracle, but IBM has recently shifted its strategy to leading with
DB2/6000 Parallel Edition rather than Oracle. Lou Gerstner has recently given
a corporate directive of ``IBM first'' which has soured a lot of its third
party relationships (Oracle and Informix, in particular). SP2 is a
reasonably well-designed MPP architecture with a good switch implementation.
Scalability on DB2/6000 Parallel Edition, Informix and Oracle has proven
pretty linear for most applications. Oracle must make use of special (IBM
developed) software called a VSD (Virtual Shared Disk) I/O device driver
to give the illusion of a shared everything disk environment on the shared
nothing hardware implementation of the SP2. The VSD chews up at least 10% of
the CPU capability in each SP2 node. Old-line IBM account representatives
will try to sell their customers on tried and true MVS systems in the new
System 390 implementation. This is a clustered SMP with the new CMOS
mainframe chips. Price-performance is much better than traditional
mainframes, but still falls short of true open systems
products. Customers do not perceive ``Open MVS'' to be an open systems player.
Strengths: Marketing, marketing, marketing! Most MPP installs.
Weaknesses: No reasonable SMP for high-end open systems market,
in-fighting between RISC/UNIX open systems specialists and S390/MVS
``old guard'', still on release 2 of UNIX system V, eroding relationships
with merchant database providers, DB2/6000 not considered ``merchant
database'' by marketplace.
Selling TurboLaser Against International Business Machines:
Emphasize the risk in adopting MPP technology where it is not required.
Since IBM has no high-end SMP, it will pitch the SP2 for almost any
open systems VLDB requirement. The TurboLaser will provide equivalent
processing power to a thirty-two node SP2 system - why would anyone
choose an MPP over an SMP given the risks and unavailability of scalable
software associated with an MPP? Reports from the Meta Group (a consulting
organization similar in stature and positioning to Gartner) repeatedly
warn customers about the dangers of adopting MPP systems when an
SMP can do the same job (with lower risk and cost). Also, IBM's relationship
to Oracle is increasingly degenerating because IBM has begun marketing
DB2/6000 Parallel Edition directly against Oracle on the SP2. Customers
who want Oracle need to be aware that the IBM-Oracle relationship is on
very unstable ground.
IBM will begin marketing DB2 4.0 on the System/390 in vigor sometime this
Summer or Fall. This is an old-line mainframe MVS (clustered SMP)
implementation and customers should be queried on their commitment to open
systems standards. "Open MVS" is nothing more than POSIX compliant APIs
implemented on top of MVS. Open systems software and a modern development
environment are severely lacking on this system. TurboLaser will have
a substantial price-performance advantage versus Systems/390. Although
they will claim a price-performance shift in moving from ECL to CMOS
technology in the chip set, the shift is not as far down on the
price-performance curve as TurboLaser because they are still chained to
a CISC (Complex Instruction Set Computer), as opposed to RISC (Reduced
Instruction Set Computer) in the Alpha chip architecture, and do not have
the same volumes as Alpha which ships in a larger array of computing products.
6. nCUBE
The nCUBE product is an MPP architecture constructed from
a custom chip set which integrates CPU and communications capability into a
single component. This custom chip set is both the strength and weakness
of the nCUBE architecture. The integration of the CPU and communications
capability allows a tight packaging of large MPP configurations and
aggressive pricing characteristics. However, the fact that a custom CPU
is used puts the nCUBE on a very undesirable technology curve. As a
start-up company, it is not possible for nCUBE to deliver chip sets
with competitive performance to the large chip manufacturers in like
time frames. The nCUBE provides a CPU that rates at a mere 7.5 VAX MIPS -
woefully underpowered in today's marketplace. The nCUBE2 will up this
processing power, but nCUBE will continue to lag large chip manufacturers
by at least three generations in chip technology. The nCUBE machine
runs the Vertex operating system on each of its MPP nodes. Despite market
positioning in the UNIX camp, Vertex is little more than a low-level
file system kernel and is inappropriate for development of general purpose
applications. nCUBE runs only the Oracle RDBMS and is primarily owned
by Larry Ellison, CEO of Oracle. Very little third party software exists
for the nCUBE machine, so customers should think of it as an Oracle server
rather than a general purpose processing engine. The Oracle MPP software
was originally developed on the nCUBE (due, in a large part, to Larry Ellison
influence) and yet this platform has been abandoned as the reference MPP
platform in favor of the IBM SP2. Andahl had an OEM agreement with nCUBE
to re-sell the MPP as a database server under the auspices of the
Explorer/2000 product line. Amdahl abandoned the product after it failed
to get adequate performance and reliability out of the nCUBE platform.
Strengths: Relationship with Oracle, packaging.
Weaknesses: Very weak CPUs, scalability and reliability issues
related primarily to glitches in Oracle, no strong reference sites,
lack of general-purpose processing capability, lack of third party tools.
Selling TurboLaser Against nCUBE:
Emphasize (1) risks of an MPP implementation where it is not needed, (2)
lack of a general purpose operating system, and (3) severely under powered
CPUs. It would probably take nearly 100 nodes in an nCUBE configuration
to match up to a high-end TurboLaser configuration. The risk in deploying
and managing a large number of nodes in an MPP configuration are not
justified when the same work can be handled in a low-risk TurboLaser
configuration. nCUBE is at a particular disadvantage in this regard
because its CPUs are custom designed and many generations behind modern
price-performance curves. Point out that Amdahl recently had an OEM
relationship with nCUBE that is terminated because the product was
unable to obtain reasonable performance or reliability characteristics
for high-end applications. Furthermore, lack of a general-purpose
operating system means no third party software and will relegate the
machine to a dedicated DBMS server - another (general purpose) machine will
need to be brought in for data preparation, data mining, and any other tasks
that require any software other than Oracle.
8. Sequent
Sequent is the most mature player in the UNIX SMP
marketplace. The Dynix operating system is quite scalable and robust
compared to most other (less mature) SMP operating system implementations.
Sequent has clustering capability with its SMPs, but has no credible
MPP strategy. They will be addressing the lack of query scalability
across SMP nodes in a clustered architecture by introducing a SCSI
connect (or something similar) between nodes so that data can be
shared at high bandwidths. This is somewhat of a kludge to work around
lack of an MPP strategy and the fact that the Intel-based SMPs are
under powered. Sequent market share is clearly declining. As a means of
addressing shrinking margins and market share in hardware sales, Sequent
has been re-positioning itself as data warehouse "consultants" in the
industry. The attempt is to build up the service revenues within the
company, but the actual expertise that Sequent has in this area is
fairly limited.
Strengths: Operating system scalability, commercial references and experience,
Oracle and Informix reference platform for SMP, market visibility.
Weaknesses: Intel chip set, hardware scalability, small company,
de-focused on NT server, still on release 3 of UNIX System V.
Selling TurboLaser Against Sequent:
Emphasize company viability and raw processing power. Sequent is a
relatively small company that is feeling a lot of competitive pressures.
They recently invested a lot of dollars and development efforts into the
NT server which has basically flopped in the marketplace. They are
still running release 3 of System V UNIX. Emphasize DEC's modern UNIX
implementation with 64-bit capability (no 2GB file limits!) versus
the antiquated Sequent Dynix implementation. Also note that raw processing
power in the Sequent is severely limited by use of the Intel chip set. An
eight CPU TurboLaser is easily more powerful than the largest configurable
Sequent SMP machine.
9. Silicon Graphics, Incorporated
Silicon Graphics now makes over half of
its revenue selling server systems. The server product line offers both
SMP and clustered SMP configurations. The SMP will configure with as many as
thirty-six 200 MHz MIPS R4400 chips, although Oracle will only scale to
between thirty and thirty-two CPUs. SGI has begun to bid Informix on some of
its high-end DSS deals, but still mainly focuses on Oracle. SGI uses
ASIC (Application Specific Integrated Circuit) technology to build one
of the fastest single bus structures in the industry at 1.2 GB/s.
Despite presence in the server market, most commercial customers
will associate SGI to Jurassic Park graphics and the high-end workstation
market. SGI has recently embarked upon an aggressive campaign in the
United States and Europe to re-position itself for data warehouse and data
mining applications. They have hired a large number of marketing and
technical people away from Hewlett-Packard and Pyramid to assist in the
transition to the commercial marketplace. The RAS (reliability, availability,
and serviceability) positioning of the SGI servers is somewhat weak - no
RAID capability, lack of hot pluggable components, etc. - so most of the
focus for this product line has been on the data warehouse environment where
RAS is less of an issue than mission critical OLTP. The focus on data
warehousing also allows leverage off of strength in graphics by emphasizing
the data visualization aspects of data mining applications. An EDS spin-off,
known as dbIntellect, uses the SGI Challenge servers almost exclusively
in their data warehouse integration product and services package. Tandem
also OEMs the SGI challenge series under the name of Integrity in the Tandem
product line for customers who want UNIX servers rather than Guardian
servers.
Strengths: Data visualization software, high bandwidth I/O subsystem,
latest MIPS chip release, aggressive marketing initiative in data mining,
EDS partnership.
Weaknesses: Lacks a total solution perspective, weak RAS story,
thought of as engineering/scientific player rather than commercial player.
Selling TurboLaser Against Silicon Graphics, Incorporated:
Emphasize total solutions perspective and RAS. DEC's expertise and
product offerings for enterprise systems integration, networking,
and overall systems management will make any customer feel much more
comfortable than SGI's "hot box" marketing strategy. Plus, TurboLaser
is a "hotter" box anyway, so SGI lacks their usual story in high-end
bus and CPU components when compared against TurboLaser. The SGI RAS
story is weak - no RAID capability, lack of hot pluggable parts, single
point-of-failure architecture, and so on. The workstation and
scientific/engineering bias in their product offerings plays poorly
for data server positioning.
10. Sun Microsystems
Sun has been slowly inching up from the networked workstation market into
data servers with its SPARC Centre 1000 and 2000 products. The SC 2000
configures up to 20 SPARC CPUs into an SMP configuration. Earlier versions
of Solaris provided poor scalability on SMP hardware, but the most recent
release allows scalability to at least fourteen CPUs. Sun competes primarily
on price. The price per megabyte on memory and disk that Sun offers,
and price per CPU, is very difficult to beat. Price per megabyte on the SPARC
Storage Arrays is well under one dollar. Almost all software in the market
will run on the Sun product line as a tier one port (if not as the reference
platform). Sun is not really set up to handle the long sales cycles
and benchmark scenarios that often result in the high-end server market.
They also do not service high-end customers very well. A joint venture
with Amdahl (similar to that struck by Cray Research Superservers) to
OEM the SPARC Center product has been put into place to address these issues.
Strengths: Third party software, high visibility, aggressive pricing.
Weaknesses: Poor customer support, medium range performance ceiling.
Selling TurboLaser Against Sun Microsystems:
Emphasize customer support and raw processing power. Sun customer support
has one of the worst reputations in the industry for high-end data server
machines. Expertise in systems engineers and the company culture is oriented
toward networked workstations rather than data servers. Sun has not yet
realized that the support requirements and skill sets to provide such support
are different in the data server environment versus the desktop. TurboLaser
also has a substantial advantage in raw processing power. Even if a
SPARC Center 2000 scaled all the way up to 20 CPUs (which it doesn't),
an eight CPU TurboLaser would easily surpass its performance capability
due to the power of the Alpha chip as compared to the SPARC chip. Sun
is a mid-range SMP competitor, whereas TurboLaser goes from mid-range all
the way up to the top of the high-end.
11. Tandem Computers
The Himalayan K10000 from Tandem has received a lot of attention
as a result of its world record breaking performance on the TPC-C benchmark.
The K10000 is an MPP architecture built from 200 MHz R4400 chips.
In reality, calling it an MPP architectureis a bit of a stretch since
the interconnect is really a combination of the old dual Dynabus (2x20MB/s)
structure for horizontal slices of the Torus with vertical connectivity
provided by 4 MB/s token rings. This is not truly a scalable interconnect,
but these nuances are sure to be addressed in a future product release. The
Tandem machine derives much of its performance advantage over competitors
from the fact that it uses a proprietary software implementation.
A tight integration of the TP Monitor, Operating System, and DBMS
on Tandem allows a tight code path in getting transactions in and
out of the system. This is more of a differentiator for OLTP than DSS
because getting in and out of the database is a larger portion of transaction
execution for OLTP. The Tandem Non-Stop SQL RDBMS, with its shared-nothing
architecture and global indexing structures, is the only RDBMS
implementation on the marketplace today that offers generalized scalability
for OLTP in an MPP environment. Tandem also offers excellent RAS capability
with its high-availability architecture and superior performance tuning
and monitoring tools.
Tandem is still learning about the DSS marketplace, but is aggressively
pursuing market share in this area. The Non-Stop SQL implementation lacks
some of the features that are important for DSS (e.g., substring, arithmetic
operators, etc.), but Tandem is being very aggressive about making these
features available upon customer demand. Tandem's major technical strength,
it's proprietary software architecture, is also its major shortcoming in
the marketplace. The Guardian operating system is completely proprietary
and lacks a modern software development environment and third party
software.
Tandem is addressing this shortcoming by development of a
"UNIX personality" that will run on top of its Non-Stop kernel so that
the advantages of UNIX can be combined with the RAS and performance
characteristics of the Non-Stop kernel. The outcome of this development is
still unclear as to what performance and RAS compromises will exist in
the UNIX personality of the Non-Stop kernel. Moreover, Tandem will need
to convince third party vendors to port their software to its UNIX
implementation. Lastly, it must convince customers that it is serious about
open systems and it committed to a UNIX direction. The jury is still
out on these points. Even in the UNIX Personality of the Non-Stop kernel,
Tandem will continue to run its proprietary Non-Stop SQL -
although (unverified) rumors exist that an Oracle port will also be
implemented. To support the investment and value-added that Tandem
provides in its proprietary software base, Tandem is priced at a premium
relative to all product offerings (with the exception of Teradata). It
is unclear how Tandem will perform its pricing in the more competitive open
systems market.
Strengths: Performance and scalability (particularly in OLTP),
overall RAS, visibility.
Weaknesses: Proprietary, poor software development environment, lack
of third party software, pricing.
Selling TurboLaser Against Tandem:
Emphasize open systems commitment and price-performance. Tandem
runs a proprietary database on a proprietary operating system with a
proprietary TP monitor (with proprietary systems management tools).
This is not a story that customers in the 1990s should be buying into.
To address this concern, Tandem will emphasize their ``UNIX personality.''
Have the customer check Tandem references that are production in the
UNIX implementation. There are none. Even if the UNIX implementation were
viable, should the customer be considering a proprietary database over
Oracle? Also, in order to support all of this proprietary software that
must be maintained in-house at Tandem, they will be very high priced as
compared to equivalent TurboLaser configurations.
Appendix I: Competitive Matrix
Competitor Strengths Weaknesses
================ =========================== =========================
AT&T GIS Teradata reference sites Price-performance
3600 Solutions focus Scalability in UNIX SMP
Intel chip set
I/O subsystem for cluster
Proprietary Teradata DBMS
Sybase Navigator performance
Cray Price-performance 16 CPU minimum configuration
CRS 6400 Solaris ABI compliance Vendor viability
Third party SW availability DSS benchmark performance
Amdahl OEM relationship Availability strategy
Scales well to 30-32 CPUs Customer support structure
OLTP benchmark performance
Data General Price-performance Marketing
Avion Series Vendor viability
Market visibility
Hewlett-Packard Powerful CPU Scalability
T500 Price-performance RISC chip at end of lifetime
Service reputation Intel relationship
Open systems commitment
Third party SW availability
IBM Marketing Poor SMP offering
SP2 MPP reference sites UNIX vs. MVS in-fighting
Third party SW availability Still on SVR2 UNIX
DB2/6000 not merchant DBMS
Relations with DBMS vendors
nCUBE Oracle relationship Weak CPUs
nCUBE2 Packaging Scalability in Oracle port
Reliability
Lack of strong references
Proprietary operating system
Lack of third party SW
Pyramid Oracle relationship I/O subsystem
Nile/RM1000 Quality of Oracle port Life after SNI acquisition?
Both SMP and MPP offerings No low-end products
Seamless path: SMP to MPP
High-end focus
Commercial references
Benchmarking expertise
Operating system scalability
Sequent Oracle relationship Intel chip set
Symmetry Series Informix relationship Small company
Commercial references Still on SVR3 UNIX
Market visibility De-focused on NT server
Operating system scalability No MPP strategy
Silicon Graphics High bandwidth I/O subsystem Lacks total solution approach
Challenge XL Latest MIPS chip integration Weak RAS story
EDS partnership Lack of commercial experience
Data mining software
Data visualization software
Sun Microsystems Price-performance Poor customer service
SC 2000 Market visibility Performance ceiling
Third party SW availability
Tandem Proven scalability Proprietary operating system
Himalaya K10000 OLTP performance Proprietary DBMS
TPC-C results Proprietary TP Monitor
Overall RAS Lack of third party SW
Market visibility Poor SW development tools
Pricing
|