Performance Improvements: Hardware vs. DB Design

Hi
I recently designed & developed a data warehouse for a client of ours.
The project has been quite a success, but they have since loaded the
database with a substantial amount of (historic) data, and have given
access to the query & reporting system (Business Objects) to many more
users. As a result, the query time on the reports is beginning to
slow.
I need to advise the client on what should be done to improve
performance again.
I'm quite familiar with what can be done on the design side of the
database (i.e. indices, aggregate tables, etc.), but I am not very
sure about the cost/benefits from a hardware perspective. As a
consultancy, the cost of our services (to build aggregate tables, etc)
would be quite expensive. It might be a better option for the client
to improve the spec of the server instead, but I am reluctant to
recommend this option without knowing that there would be guaranteed
benefits. In other words, if they were to double the processing power,
would the query time effectively halve? What about memory - i.e. would
more memory help with multiple-users, but not affect the speed of the
reports? Is 4GB the max memory that Windows 2000 can have?
[I can't remember the exact spec, but the server's got something like
2 Xeon dual-processors, 4GB RAM & Win2k.]
Can anyone recommend a study or white-paper on the performance
improvements in relation to hardware upgrades, or something similar?
Otherwise, what other options do I have? Am I overlooking something
like two load-balanced servers?
Thanks for any help / recommendations!
Sean W.Sean wrote:
> Hi
> I recently designed & developed a data warehouse for a client of ours.
> The project has been quite a success, but they have since loaded the
> database with a substantial amount of (historic) data, and have given
> access to the query & reporting system (Business Objects) to many more
> users. As a result, the query time on the reports is beginning to
> slow.
> I need to advise the client on what should be done to improve
> performance again.
> I'm quite familiar with what can be done on the design side of the
> database (i.e. indices, aggregate tables, etc.), but I am not very
> sure about the cost/benefits from a hardware perspective. As a
> consultancy, the cost of our services (to build aggregate tables, etc)
> would be quite expensive. It might be a better option for the client
> to improve the spec of the server instead, but I am reluctant to
> recommend this option without knowing that there would be guaranteed
> benefits. In other words, if they were to double the processing power,
> would the query time effectively halve? What about memory - i.e. would
> more memory help with multiple-users, but not affect the speed of the
> reports? Is 4GB the max memory that Windows 2000 can have?
> [I can't remember the exact spec, but the server's got something like
> 2 Xeon dual-processors, 4GB RAM & Win2k.]
> Can anyone recommend a study or white-paper on the performance
> improvements in relation to hardware upgrades, or something similar?
> Otherwise, what other options do I have? Am I overlooking something
> like two load-balanced servers?
> Thanks for any help / recommendations!
> Sean W.
Unfortunately, I think many companies are inclined to add more server
without first trying to tune the database. You've gone ahead and
designed a database for the customer and added the necessary RI and
indexes. But now that you see the SQL executed, someone needs to tune
the queries, revise the indexes if necessary, and verify the reporting
system is executing SQL in an efficient way: For example, is it using
stored procedures? Do users have Ad Hoc access to the database to query
whatever they want? Does the reporting tool bring back large result sets
and filter on the client PC? Do the queries use a READ UNCOMMITTED or
NOLOCK query option to keep locks to a minimum?
My point is that without performance tuning the queries, it's really
impossible to know whether adding more hardware will give you anything
but a temporary reprieve of poor performance.
When I hear about problems like this, I like to remember an old client
that had a couple hundred people hitting a SQL Server database which ran
on a dual-CPU 486 with 84 MB of RAM. I know times have changed, but when
I hear about performance problems on quad-Xeon systems with 3GB RAM (as
an example), I get concerned that a company is considering throwing
another $150K at a new server.
Ad Hoc reporting systems are notoriously bad on performance, especially
if users don't run canned reports and can throw what SQL the end-user
reporting tool can generate at the server. The amount of data queried
can easily move old data out of cache and force the next query to read
that data from disk, which is really slow. Adding more memory is an
option if plan reuse is the issue. And since memory is cheap, it can't
hurt to add all you can. Adding more hardware will help, but may not
really be a long term solution.
However, nothing beats being able to aptly use an index to generate a
result set. Table/Clustered Index Scans are a killer on big tables.
You may want to consider placing query limits on ad hoc queries using
"SET QUERY_GOVERNOR_COST_LIMIT ". You can also use Profiler to track
those queries cosuming excessive CPU and see if the database or the
queries can be tuned.
--
David Gugick
Imceda Software
www.imceda.com|||With Business Objects it is worth asking the users to send you copies of
their reports. You can then see if indices need creating on various
tables etc.
Adrian
David G. wrote:
> Sean wrote:
>>Hi
>>I recently designed & developed a data warehouse for a client of ours.
>>The project has been quite a success, but they have since loaded the
>>database with a substantial amount of (historic) data, and have given
>>access to the query & reporting system (Business Objects) to many more
>>users. As a result, the query time on the reports is beginning to
>>slow.
>>I need to advise the client on what should be done to improve
>>performance again.
>>I'm quite familiar with what can be done on the design side of the
>>database (i.e. indices, aggregate tables, etc.), but I am not very
>>sure about the cost/benefits from a hardware perspective. As a
>>consultancy, the cost of our services (to build aggregate tables, etc)
>>would be quite expensive. It might be a better option for the client
>>to improve the spec of the server instead, but I am reluctant to
>>recommend this option without knowing that there would be guaranteed
>>benefits. In other words, if they were to double the processing power,
>>would the query time effectively halve? What about memory - i.e. would
>>more memory help with multiple-users, but not affect the speed of the
>>reports? Is 4GB the max memory that Windows 2000 can have?
>>[I can't remember the exact spec, but the server's got something like
>>2 Xeon dual-processors, 4GB RAM & Win2k.]
>>Can anyone recommend a study or white-paper on the performance
>>improvements in relation to hardware upgrades, or something similar?
>>Otherwise, what other options do I have? Am I overlooking something
>>like two load-balanced servers?
>>Thanks for any help / recommendations!
>>Sean W.
>
> Unfortunately, I think many companies are inclined to add more server
> without first trying to tune the database. You've gone ahead and
> designed a database for the customer and added the necessary RI and
> indexes. But now that you see the SQL executed, someone needs to tune
> the queries, revise the indexes if necessary, and verify the reporting
> system is executing SQL in an efficient way: For example, is it using
> stored procedures? Do users have Ad Hoc access to the database to query
> whatever they want? Does the reporting tool bring back large result sets
> and filter on the client PC? Do the queries use a READ UNCOMMITTED or
> NOLOCK query option to keep locks to a minimum?
> My point is that without performance tuning the queries, it's really
> impossible to know whether adding more hardware will give you anything
> but a temporary reprieve of poor performance.
> When I hear about problems like this, I like to remember an old client
> that had a couple hundred people hitting a SQL Server database which ran
> on a dual-CPU 486 with 84 MB of RAM. I know times have changed, but when
> I hear about performance problems on quad-Xeon systems with 3GB RAM (as
> an example), I get concerned that a company is considering throwing
> another $150K at a new server.
> Ad Hoc reporting systems are notoriously bad on performance, especially
> if users don't run canned reports and can throw what SQL the end-user
> reporting tool can generate at the server. The amount of data queried
> can easily move old data out of cache and force the next query to read
> that data from disk, which is really slow. Adding more memory is an
> option if plan reuse is the issue. And since memory is cheap, it can't
> hurt to add all you can. Adding more hardware will help, but may not
> really be a long term solution.
> However, nothing beats being able to aptly use an index to generate a
> result set. Table/Clustered Index Scans are a killer on big tables.
> You may want to consider placing query limits on ad hoc queries using
> "SET QUERY_GOVERNOR_COST_LIMIT ". You can also use Profiler to track
> those queries cosuming excessive CPU and see if the database or the
> queries can be tuned.
>|||Adrian Edwards wrote:
> With Business Objects it is worth asking the users to send you copies
> of their reports. You can then see if indices need creating on various
> tables etc.
> Adrian
>
Good point. If you can get your hands on any canned reports or at least
have a process for quickly rolling out new or updated reports, you may
eliminate users creating queries that don't perform well.
You might want to consider having new reports created against a test
server before moving them into production. Not giving users the rights
to create reports on the production data may be difficult to manage.
Managers generally want fast results. If you can't do that, at least
introduce a review process so there's never a time when too many reports
are created without review.
You can do this passively using Profiler or a server-side trace. You
could trap the SQL:BatchCompleted and RPC:Completed events for CPU over
a certain level (say 1,000ms) and review those SQL Statements
periodically.
David Gugick
Imceda Software
www.imceda.com