| [Digital Equipment Corporation]
[Enterprise solutions initiative]
[redbar]
SUBJECT: Responding to Sun's "Supercaching for Very Large Databases" Paper
Sun has developed a whitepaper -- "SuperCaching for Very Large Databases"
-- which describes a proprietary 32-bit software scheme to access greater
than 4GB's of physical memory. In this paper, Sun minimizes Digital's
64-bit VLM and VLDB capabilities as being of little consequence.
Sun is using this whitepaper to help Sun sales people sell a proprietary
scheme to enable 32-bit Solaris to access large (greater than 4GB) physical
memory against Digital's full 64-bit VLM/VLDB capabilities.
The bottom line is that Sun's proprietary scheme is more difficult to
deploy, does not deliver the application and database performance
advantages of Digital's full 64-bit solution, and will require multiple
customer application and database migrations over the next several years.
Because we have seen this whitepaper attached to several Sun RFP's where
Digital has been in the running, this memo is being written and is intended
to:
- ALERT YOU to the fact that this is how Sun is attempting to sell
against a clear Digital advantage.
- PROVIDE YOU with an explanation of Sun's 32-bit SuperCaching
capabilities.
- ARTICULATE the advantages of Digital's full 64-bit VLM/VLDB
capabilities as contrasted with Sun's 32-bit SuperCaching
capabilities.
Please share this memo with anyone you think will benefit from this
information. This memo will also be posted on the turbolaser-inside
intranet site, at:
http://decpos2.eng.pko.dec.com/turbolaser-inside ... in the Competitive
Profiles section
By intent, we are NOT publishing the Sun whitepaper as we do not wish to
proliferate Sun's anti-Digital marketing collateral. However, if you feel
you need to have a copy of this paper to help combat a Sun tactic, please
send me a note at lando::warshawsky, or [email protected] and I
will get you a copy.
In their paper, Sun states that ...
"The industry is vibrating with hubbub about caching entire databases, a
technique referred to as supercaching. Some vendors have been stressing the
use of 64-bit virtual addressing to implement super-caches. This paper
describes some of the assertions and outlines some equally effective
alternative methods, including configurations that can be implemented on
32-bit Solaris 2.x. We provide some benchmark results showing the
effectiveness of the technique, and discuss some of the advantages of the
various proposed methods. Lastly we consider the applicability of the
supercaching technique, with special attention on practical and economic
issues."
Essentially, Sun defines supercaching simply as the ability to cache very
large; i.e. more than 4GB, data stores in memory. Sun further states that
"although it is intuitively obvious how to cache very large databases with
64-bit, this feature is NOT required to achieve these (spectacular
performance) results ..." [ed.: Sun's wording -- yes, Digital HAS
demonstrated spectacular performance gains with VLM]. Sun goes on to say
that they have developed an alternate (to true 64-bit VLM technology)
proprietary scheme for caching large Oracle databases on their 32-bit
systems. Sun makes no mention of support for other databases or
applications.
Certainly 32-bit Solaris systems can structurally address large physical
memories. Even 16-bit systems can do this, if you don't mind the extra
overhead and inconvenience of software "hacks" and kludges you have to put
up with. But this is non-standard.
The way 32-bit Solaris 2.x handles large physical memory for Oracle
databases is through a proprietary combination of a small Oracle SGA and
the Solaris UFS cache, whereas on a true 64-bit system such as Digital's
TurboLaser, the Oracle SGA is large and is able to embrace ALL of physical
memory without the creation of special intermediate holding areas or
caches.
Assuming a large UFS cache, the extra overhead in CPU due to the
second-level cache scheme is claimed by Sun to be in the range of 5%-8%
over the single-level cache (as implemented in a true 64-bit
implementation). However, in real-world transaction processing, when
transactions are committed, log files are written directly to disks
bypassing the UFS. Thus, with small SGA buffers, performance will be
further degraded and this degradation can only be minimized by extra tuning
exercises. In addition, many other Oracle operations bypass the UNIX file
systems and transfer data directly between SGA and disks. Therefore, with a
small SGA, there will be slowdowns beyond the 5%-8% range, perhaps as much
as 10-15%, depending on the type of workloads.
Additionally, since all Oracle writes are "write through" the file system
buffer, a small SGA (as implemented by Sun) will trigger frequent disk
writes and therefore a large UFS file system buffer will not be very
helpful here. A large SGA (as implemented by Digital) will defer immediate
I/Os, resulting in better OLTP performance.
In general, Oracle management of the database buffer is more efficient than
the UNIX management of the database in a file system buffer. The large SGA
database buffer implemented by Digital allows for selective caching of
database objects, such as specific large tables, whereas this is
impractical in Sun's implementation. Also, the two levels of caching as
implemented by Sun tends to duplicate the data leading to some waste of
memory. And finally, the performance improvement of using raw partitions
beyond the file system (as in Digital's implementation) is well established
and is generally in the range of 10 to 15%.
Another point worth considering. If a computer is running database
applications concurrently with non-database workloads, SUN will have to
come up with yet another set of kludges to maintain system performance,
such as keeping non-database workloads resident in memory, while still
assuring that enough UFS cache is still available. Since the UNIX file
system buffers are database independent, a less important non-database user
may overwrite all cached database data. If there are multiple instances
running in the system, an efficient cache allocation becomes very difficult
with file system buffers, requiring further special techniques.
And finally, as the rest of industry is moving toward full 64-bit system
solutions, the inconvenience (for customers who buy into Sun's scheme) of
coming up with special tricks to stay with a 32-bit system will rapidly
deteriorate into a major headache. How will a customer solution based on
this scheme migrate with the multiple planned Solaris release's between and
now and when Solaris becomes full 64-bit? Will backwards compatibility for
this scheme be a priority for the Solaris development group? If history is
a guide -- NOT!
... and what about Oracle? With Oracle 8 just around the corner, how will
an Oracle 7 "supercache" solution migrate? Is Sun prepared to underwrite
this cost for their customers or will they merely inform their customers of
the price and penalties later?
In summary, the penalties and disadvantages for Sun's supercaching scheme
are:
1. Performance: Sun's scheme is based on a combination of a small Oracle
SGA and Sun UFS buffer cache, which translates into additional system
overhead resulting in less than optimal performance. Under ideal
conditions, Sun estimates 5-8% performance degradation as compared to a
full 64-bit implementation, such as Digital's. Digital believes in
practice, the performance degradation could be up to twice as much
depending on the workload and other concurrent applications.
CHALLENGE Sun to produce audited TPC-C numbers beyond Digital's reported
results!
2. Proprietary: Although Sun's supercaching scheme can be made to work, the
scheme is proprietary. How will a customer solution based on this
implementation migrate with the multiple planned Solaris releases between
now and when Solaris becomes full 64-bit? Will backwards compatibility for
this scheme be a priority for the Solaris development group (-- my guess is
not)?. And, by the way -- will Sun be disavowing this approach once SUN has
full 64-bit capabilities? You can bet they will!
CHALLENGE Sun to tell you how their software workaround scheme will evolve
as Solaris evolves towards a full 64-bit capability.
3. Oracle only: Sun's scheme is specific to Oracle. With Oracle 8 just
around the corner, will Oracle 7 implementations work with Oracle 8? What
about other merchant databases? Digital offers true VLM/VLDB capabilities
with investment protection and without business interruption for all
merchant databases.
CHALLENGE Sun to tell you what their supercaching strategy is for all other
merchant databases ... and for Oracle8.
4. Database Management: Scheme will likely require customer changes in db
structures and management as Solaris evolves towards delivering full 64-bit
capabilities (est. 1999).
CHALLENGE Sun to underwrite the cost of writing and maintaining this
proprietary software scheme throughout various release of Oracle and
Solaris.
Bottom line: Sun's proprietary scheme is more difficult to deploy, does not
deliver the application and database performance advantages of Digital's
full 64-bit solution, and will require multiple customer application and
database migrations over the next several years.
****
Thanks to Kuk Chung, Manager, Database Technology Center, Palo Alto, John
Henning, CSD Performance Engineering Group, John Bennett and Dave Poole,
Enterprise Servers, for their review and inputs to this memo.
[redbar]
[Local home] [navbar]
Updated: 30 December 1996
Legal
|
| [First! for the Web]
Data Warehousing Expands -- Vendors ready tools and databases for booming
market
---------------------------------------------------------------------------
Date: Tuesday, February 25, 1997
Source: Information Week
---------------------------------------------------------------------------
Information Week via Individual Inc. : The data-warehousing boom shows no
signs of a slowdown. The market is expected to grow 50% this year, to $5.7
billion, according to the Meta Group, a research firm in Stamford, Conn.
Database and tools vendors are readying products and alliances to get a
bigger piece of the pie. IBM is expected to expand its data-analysis
product line this week via deals with Sybase Inc. and Arbor Software Corp.
IBM's RS/6000 division is teaming with Sybase in Emeryville, Calif., to
sell data marts based on Sybase IQ, a database that specializes in
high-speed analysis, and IBM's Unix- based RS/6000 servers. Prices will
begin at $180,000 for a four-processor RS/6000 with a five-user license.
IBM's Software Solutions group is expected to broaden an existing pact with
Arbor Software, in Sunnyvale, Calif., to integrate components of Arbor's
Essbase database into IBM's DB2. This will give DB2 users additional data-
management and manipulation functions such as data retrieval, storage, and
multidimensional analysis, sources say. "I expect this to counter Oracle's
Oracle7 and Oracle Express combo," says Frank Gillett, an analyst with the
Hurwitz Group Inc., a Newton, Mass., consulting firm.
Also this week, SAS Institute Inc. will announce a metadata management
tool, the SAS Warehouse Administrator, a central console for tracking data
as it moves from a database to a data warehouse. "This represents our first
attempt at coming up with a common metadata platform and repository," says
Jim Davis, program manager of data warehousing with SAS, in Cary, N.C.
Other tools are expected at DCI's Data Warehouse Conference this week in
Orlando, Fla. (see p. 76)
Meanwhile, corporate users are leveraging data warehousing in creative
ways. Credit-card giant Visa International is launching severalinitiatives
to help its member banks be more profitable. "There's almost nothing in the
company that won't be touched by where we're going with data warehouses,"
says Deborah McWhinney, executive VP with Visa in San Francisco.
Visa is using data warehouses on several fronts, including target
marketing, fraud and risk analysis, and financial reporting. It's also
developing predictive modeling software to identify card holders who
present a high risk for bankruptcy. The projects are part of Visa's
Information Initiative, a strategy for applying IT to meet several business
objectives.
Says McWhinney, "Information is becoming a much bigger piece of what we
offer our members."
Copyright 1997 CMP Media Inc.
<<Information Week -- 02-24-97, p. 16>>
---------------------------------------------------------------------------
[Feedback] [Index] [Search] [Customize] [Individual, Inc.]
[Main] [Up] [< Prev Story] [Next Story >] [Calendar]
---------------------------------------------------------------------------
[Individual, Inc.]
|
| [First! for the Web]
Winter Corporation identifies world's largest databases
---------------------------------------------------------------------------
Date: Tuesday, April 15, 1997
Source: Business Wire
---------------------------------------------------------------------------
BOSTON--(BUSINESS WIRE) via Individual Inc. -- Technology pioneers who
build and manage leading IT systems are honored at VLDB Summit Conference
Winter Corporation, a consulting and research firm specializing in large
database technology, announced the Grand Prize winners of its 1997 survey
program at the VLDB (Very Large Database) Summit Conference in Chicago.
Findings from the annual program help VLDB users maximize the technical
efficiency and business value of these key corporate assets. Survey results
are also of value to industry vendors who seek to prioritize internal
resources to meet the needs of the VLDB user community.
"This year, we received surveys from fifteen different countries, including
the Middle East, Pacific Rim and Eastern Europe," announced Richard Winter,
President of Winter Corporation and a leading expert on designing and
building large databases for the global enterprise. "The Grand Prize
winners we are announcing today represent a cross-section of industries,
from online information services and retail business to transportation and
telecommunications. The enthusiastic and wide-ranging response to our
program shows that organizations - around the world and across all
industries - are committed to improving their use of corporate data to make
critical business decisions, gain competitive advantages and reach
strategic objectives."
The databases maintained by the participants in the VLDB Survey Program
touch our lives every day. For example, they pay Social Security, locate
shipments in transit, indicate when to restock shelves in our local retail
stores, and ascertain which health care services to assign to particular
individuals.
For the third year, the VLDB Survey Program is being conducted in
partnership with Database Programming and Design magazine. David Stodder,
editor-in-chief of the publication and chairman of the VLDB Summit
Conference, reports that findings from the survey will be featured in a
special issue to be published later this year.
In the 1997 program, nine companies, including three multiple winners,
earned first-place honors in a range of database categories. Top contenders
were chosen based on three primary criteria: largest database size
(including user data, summaries, aggregates and indexes, but excluding
freespace, RAID overhead or mirroring), maximum number of concurrent users
and most rows, records or objects. Winners were selected for both decision
support and transaction processing systems. To measure the shift towards
open systems, UNIX environments were differentiated from traditional
operating platforms.
Wal-Mart Stores, Inc., Bentonville, AR, takes the most laurels in this
year's program. Its decision support system, employing NCR Teradata DBMS
and NCR WorldMark hardware, outpaced the other UNIX entries in all three
criteria: a mammoth 2.4 terabyte database, 650 concurrent users, and 20
billion rows. This Grand Prize winning database is undergoing voracious
growth, having doubled in size over the past year and projecting a similar
gain over the next twelve months.
Knight Ridder Information, Mountain View, CA, is also a multiple Grand
Prize winner in the '97 program. Its blue ribbon text search database,
DIALOG, weighs in at seven terabytes, earning top honors as the largest
decision support database in any operating environment. The system also
takes the prize for the most rows/records, 100 billion, in this category.
DIALOG combines a proprietary database management system and Hitachi Data
Systems hardware and processes an impressive 700,000 queries per day.
Another multiple Grand Prize winner in the program is Amadeus Data
Processing GmbH and Co., Erding, Germany. A transaction processing system
running on UNIX, the reservation travel system achieves distinction for
managing 800 gigabytes of data and for supporting 1,500 concurrent users.
The system is implemented in INFORMIX ODS 7.2 running on a cluster of DEC
Alpha systems. Headquartered in Germany, the system also processes
reservation requests from branch offices in two other countries, France and
Spain.
In the competition to support the most concurrent users, KMart Corporation,
Troy, MI, heads the field of UNIX-based decision support systems. The
system allows 2,500 users to connect at the same time, running planning,
forecasting and sales tracking queries. The database is implemented in a
pure NCR environment: Teradata DBMS and a system 3600. The system allows
KMart to update its inventory on a daily basis, engendering a more accurate
in-stock position.
Dual entries from United Parcel Service, Mahwah, NJ. share the winners'
circle for the largest transaction processing database in any environment.
Each system contains about three and one/half terabytes of data and
utilizes IBM's DB2 running on IBM 9021 mainframes. One database monitors
packages as they navigate through UPS centers while the other records
delivery information. UPS has plans to expand its internal systems to
provide in-transit package information to customers.
Outdistancing the competition for the number of rows in a transaction
processing system, in any environment, is TELSTRA, a telecommunications
company based in Melbourne, Australia. A Grand Prize goes to the company
for its customer billing system, which consists of 50 billion rows of data.
The main component of the DBMS is IBM DB2, in a cluster of three IBM 9021
systems. The transaction processing database has tripled in size over the
past twelve months and is projected to grow by 50% in the next year.
State Street Bank, Quincy, MA, secures the gold ring for most concurrent
users among transaction processing systems operating in any environment.
The award winning database supports 20,000 simultaneous users performing
trading and custody activities as well as internal support functions. State
Street Bank uses CA-Datacom DBMS aboard five IBM 9000 machines. The
prodigious system supports a total of 40,000 online users, including other
banks besides State Street and their customers as well.
Top honors go to The Handleman Company, Troy, MI, for most rows in a
transaction processing, UNIX-based environment. The system employs an
ORACLE DBMS and is hosted on two Siemens-Pyramid Nile 150 machines. Data
resides in a total of 1.3 billion rows in the database, which is used for
contribution analysis and for automatic replenishment of audio and video
inventory items.
A new category was established in this year's program for federated
databases, which are virtual databases containing independently operating
databases that own their own data. National Processing Company, Louisville,
KY, finished with the top decision support system in this category. The
company maintains an audit trail system that is spread over 18 different
platforms in four locations. The UNIX-based system holds 822 gigabytes and
is used to capture and retrieve retail financial transactions. The primary
DBMS component of the database is ORACLE running under Sun SPARC, HP 9000
and IBM RS/6000 servers.
Winter Corporation and the VLDB Survey Program
Winter Corporation is an international consulting practice that offers
expertise in data warehouses, operational databases, enterprise
architectures and strategic information management decisions. The company
is unusually adept at assuring that critical technical issues - such as
database performance, scalability, availability and data quality - are
resolved in accordance with the client's business requirements.
The VLDB Survey Program is the principal data gathering mechanism of the
Winter Research and Recognition Program. The service examines a wide
variety of VLDB-related products, technical issues and organizational
approaches to create an extensive knowledge-base about very large
databases. "By illuminating best practices in VLDB management," comments
Kathy Auerbach, Research Program Manager, "the research enables users to
increase the value of their database investments and better capitalize on
fast-changing market opportunities."
All participants in the program receive the Members Report, a compilation
of selected survey findings. The report is also available to the public for
a fee. Winter Corporation offers other services related to the survey
program. These include assisting users in strategic and implementation
decisions or helping vendors with market analysis, product strategy,
performance and benchmark strategy, positioning and technical market
communications.
CONTACT: Fifth Business, Inc. | John Donnelly, 508/620-4712
---------------------------------------------------------------------------
[Feedback] [Index] [Search] [Customize] [Individual, Inc.]
[Main] [Up] [< Prev Story] [Next Story >] [Calendar]
---------------------------------------------------------------------------
[Individual, Inc.]
|