Personal Data Marts
Content
Supplement Your Microsoft Business Intelligence Strategy with the Fast Performance and Excellent ROI of Personal ata Marts
!im Peterson How fast is fast enough for OLAP browsing? OLAP is at its finest when the data changes in less than a second. The user clicks and the data changes instantaneously. This is the kind of performance that allows a user to really explore an OLAP cube drilling up! drilling down! trying one filter! adding an additional filter! switching the dimension displayed on the columns! etc. "igel Pendse wrote an article called Performance Matters for !he O"#P Report #www.olapreport.com$ in August! %&&'. (n it he says the following about OLAP tools) Slow query performance has been consistently the most serious product-related reported problem, and for the last few years it has been the single most often complained of problem. According to data gathered by the OLAP *eport! industry+wide median OLAP ,uery time a-eraged between ' and '.. seconds in %&&/.
Median input data volume and median query performance trends 0sed with permission of The OLAP *eport
1icrosoft 23L 2er-er %&&. Analysis 2er-ices pro-ides excellent OLAP browsing speed. 4ut most organi5ations using Analysis 2er-ices could benefit in certain situations with the performance boost pro-ided by personal data marts. (f your OLAP ,uery response time is o-er sixty seconds! you may be able to bring it down to fi-e seconds. (f you now ha-e a response time of ten seconds! you may be able to reduce it to one second. 6ith faster speed your OLAP cubes will pro-ide more actionable business intelligence because your business users will be willing to examine the data from more perspecti-es.
What Are Personal Data Marts and Why a!en"t # %efore&
eard About $hem
A personal data mart is a set of one or more customi5ed local cube files that ha-e been created for a specific user7s ,ueries. 8ach local cube contains a small portion of the data a-ailable in the Analysis 2er-er cube. 4ecause they are small! they can be -ery fast. 9or example! a company could ha-e a large 2ales cube! with data for :!&&& 2ales *epresentati-es for the past . years. The 2ales *epresentati-es examine the 2ales cube e-ery day. They spend most of their time analy5ing their own sales for the most recent month. A local cube containing only the data for one 2ales *epresentati-e for one month will ha-e only :;/&!&&&th of the data in the Analysis 2er-er cube. 0nused members from the Product dimension! the <ustomer dimension! and other dimensions can be remo-ed from the local cube. This small cube will browse much more ,uickly
than the Analysis 2er-er cube! especially when calculations are being e-aluated. This fast speed can be achie-ed while still keeping all of the cube7s dimensions! attributes! measures! and calculated members. 4esides impro-ing browsing speed! the local cube files also pro-ide the flexibility for the 2ales *epresentati-e to look at data when not connected to the corporate network. Local cube files ha-e been a part of the 1icrosoft 4( toolset since :==> with the release of 23L 2er-er '. Their purpose has been to pro-ide users access to OLAP data while they are offline. The use of local cubes to impro-e cube browsing speed has often not been recogni5ed. 6hile 1icrosoft has always pro-ided the ability to create local cube files! it has often been difficult to find tools to automate the repeated creation of customi5ed local cubes. 6ith the release of 23L 2er-er %&&.! the ability to design local cube files has been impro-ed. The Analysis 2er-ices 2cripting Language #A22L$ gi-es the user full control o-er the creation of the local cube. 1ost users! howe-er! do not ha-e access to a tool that creates local cubes using A22L. Local cube creation tools usually use the <reate ?lobal <ube statement! which gi-es far fewer local cube creation options. A cube created with <reate ?lobal <ube cannot be optimi5ed for browsing speed in the same way as a cube created with A22L. The result is that there are organi5ations with a lot of experience in using 1icrosoft 4usiness (ntelligence! but little experience in using local cube files. 1any of these organi5ations could achie-e a dramatic impro-ement in their cube browsing speed by switching to personal data marts for their OLAP data deli-ery or by using personal data marts for at least some of their more challenging cube browsing situations.
ow 'an # (now #f My )*AP Performance Will #mpro!e With Personal Data Marts&
The only way to know for sure! of course! is to try it. Howe-er! we ha-e found that personal data marts can be particularly useful in the following situations. $% !he cu&es ha'e one or more large( flat dimensions . 2ometimes dimensions are designed without hierarchies or with hierarchies ha-ing -ery few le-els. One example of this is the 2ales Order dimension in the Ad-enture 6orks sample database that is distributed with Analysis 2er-ices. The Analysis 2er-er cannot build effecti-e aggregations for a large! flat dimension. (f you design a local cube that remo-es a significant portion of a large! flat dimension #such as limiting the 2ales Order dimension to the sales for a single 2ales *epresentati-e$! you will greatly impro-e browsing speed when the user is looking at that dimension.
)% !he cu&es use complex calculated mem&ers% <alculations can slow browsing speed. 4ecause there is less data in a local cube! the calculations can be e-aluated more ,uickly. The impro-ed performance is most e-ident in a calculation that has to e-aluate a large number of null cells. 4ecause local cubes can ha-e all unused dimension members remo-ed! the number of null cells can be exponentially reduced. *% Each indi'idual user is using only a small portion of the cu&e . Personal data marts are especially effecti-e at impro-ing browsing speed when each user typically looks at a small subset of the cube7s data like the pre-ious example of a 2ales *epresentati-e looking at the most recent month7s data. +% Some users only want to see an odd su&set of the cu&e% ( ha-e often seen situations where users want to see an oddly shaped subset of the data such as all sales from *egion :! together with the sales from two cities in *egion %! excluding sales initiated by the corporate head,uarters staff! and only including sales to customers who ha-en7t purchased anything in the pre-ious two years. A good OLAP client tool will allow a user to see this subset of the data! but it is much easier and there can be much better performance if the user starts browsing with this exact set of data. ,% -u&e &rowsing speed is ade.uate for some &ut not all users( or is ade.uate at certain times &ut not other times% Analysis 2er-er cubes can be effecti-ely optimi5ed for the typical user! but there may be certain users who are browsing the cubes in unusual ways that greatly increase ,uery response time. 6hen one user issues a long ,uery! the ,uery response time can be slowed for e-eryone using that Analysis 2er-er. (n these situations! it can be -ery helpful to create personal data marts for the users that ha-e the lengthy ,ueries. 2peciali5ed local cubes can be designed to impro-e the browsing speed for these users. 4ecause the Analysis 2er-er no longer has to e-aluate these complex ,ueries! the browsing speed for all the other users can be impro-ed and will be more consistent.
'an # Achie!e the Same *e!el of Performance Without +sing Personal Data Marts&
(n some situations! you can match the performance of personal data marts by using other strategies. (n many cases! howe-er! these other strategies are inade,uate or may be more expensi-e to implement. $% #dd MO"#P aggregations% The Analysis 2er-er gains a lot of its ability to ,uickly respond to ,ueries through the use of aggregations. (f you ha-en7t done so already! create 1OLAP aggregations for your cubes. Aggregations will not! howe-er! increase the browsing speed in large! flat dimensions. They may or may not be helpful in impro-ing the speed of calculated members.
)% Set proper attri&ute relationships in hierarchies . 6hen using the Analysis 2er-er from 23L 2er-er %&&.! aggregations cannot be created for hierarchies unless relationships are set for the attributes used to create the le-els. *% #dd partitions to the cu&e% 6hen used with appropriate slicing in the ,ueries! partitioning can greatly impro-e cube browsing speed. (f there is one single dimension that is causing browsing speed problems! adding partitions along that dimension can be -ery effecti-e. Personal data marts! though! are more effecti-e when different users need a cube optimi5ed along different dimensions. (t7s -ery difficult to take e-ery indi-idual7s situation into account when designing partitions for an Analysis 2er-er cube. +% Optimi/e calculated mem&ers% There are many strategies for optimi5ing calculated members which! in some cases! can dramatically impro-e browsing speed. ,% 0se more powerful hardware% The Analysis 2er-er performs better with a /@+bit ser-er! multiple processors! and a large amount of memory. 1% !each users how to a'oid pro&lem &rowsing areas . 1any Analysis 2er-er cubes will perform well as long as the user doesn7t pick the wrong combinations of attributes. (n many! but not all! situations! slow performance can be a-oided by picking appropriate slicers before attributes are placed on the columns and rows. 2% Simplify the #nalysis Ser'er cu&es. Look at the list of situations where personal data marts can impro-e performance and modify the Analysis 2er-er cube so that these conditions are remo-ed) • • • • • 8liminate large dimensions *emo-e the lowest le-el of detail in large dimensions Add le-els in large! flat dimensions *emo-e calculated members if they are slowing cube browsing *emo-e seldom used attributes
(n summary! all these strategies can speed OLAP browsing. 2ome of them will make a big difference with -ery little effort. 4ut in many cases! the easiest and cheapest option is to gi-e your users personal data marts. There are many excellent resources a-ailable for optimi5ing Analysis 2er-er cubes. 2ee the end of this article for suggestions on papers! web sites! and blogs.
Personal Data Mart ,)#
(n considering the *O( of a personal data mart! there are three primary courses of action) :. 4e satisfied with cubes that ha-e slow browsing speed %. 0se one or more of the other strategies to impro-e cube browsing speed
A. 0se personal data marts to impro-e cube browsing speed 8ach of these options has a cost associated with it.
-ourse of #ction 3$ 4 Be Satisfied 5hen loo6ing at the costs of slow &rowsing speed( consider the following7
$% 5hat is the o'erall 'alue of your Microsoft Business Intelligence system8 How much did your 4( system cost? 6hat benefit is your organi5ation recei-ing from the cubes? 6hat additional benefits are a-ailable through more effecti-e use of the cubes? )% Your cu&es will not gi'e a good ROI unless they wor6 well for your users% 6hate-er benefit you recei-e from the cubes will be greatly enhanced if they can be browsed ,uickly. 9ast browsing cubes are an essential part of a 4usiness (ntelligence system. (t is unwise to spend a lot of money building a 4( system without also achie-ing the best possible cube browsing speed. *% 5hat speed is too slow8 "igel Pendse! in an article entitled What Is OLAP?, has said that a defining characteristic of OLAP is fast browsing response time where simple ,ueries take one second! most responses are returned in about . seconds! and -ery few responses take more than %& seconds. ( think this should be considered to be a minimum standard. 0sers #and organi5ations$ will gain much more benefit from OLAP if the typical ,uery is under : second and -ery few ,ueries exceed . seconds. 6ith this le-el of performance! users feel that they7re getting their ,uestions answered immediately.
-ourse of #ction 3) 4 Other Strategies
The cost is ,uite -aried for the other performance+enhancing strategies) $% Some strategies are 'ery cost9effecti'e% (f you can achie-e excellent cube browsing speed by adding aggregations! setting attribute relations! adding partitions! or by optimi5ing calculated members these would usually be the best options. 8-en if you ha-e to hire a consultant to help you do this! the benefit should be worth the cost. )% Some strategies reduce the 'alue of the cu&es% (f you ha-e to remo-e dimensions! attributes! le-els! or calculations you need to consider the cost of gi-ing up some of the cube7s functionality. Adding le-els to a hierarchy can also make a cube harder to use. The option of teaching users to a-oid certain types of browsing has a -ery high cost because it restricts OLAP browsing to the people who ha-e learned how to do it! rather than ha-ing the benefits of OLAP insight a-ailable to e-eryone in the organi5ation. *% Some strategies can &e expensi'e% The impro-ed hardware option has a definite cost attached to it. And though better hardware can be -ery helpful! by itself hardware may not bring optimal performance.
-ourse of #ction 3* 4 Personal
ata Marts
6hen using the personal data mart strategy! you need to calculate the costs of doing these tasks) $% esigning and creating the local cu&es% (t is fairly easy to create local cube files in 23L 2er-er %&&&. (n 23L 2er-er %&&.! it is usually necessary to use the Analysis 2er-ices 2cripting Language #A22L$ to create optimi5ed local cubes. The A22L can be ,uite difficult to manipulate. Our company has de-eloped a product! <ube2lice! which uses A22L to create customi5ed local cube files that can be optimi5ed for browsing performance. )% istri&uting the local cu&e files. Local cube files need to be recreated when new data is a-ailable. These new local cube files need to be distributed to the users on a regular basis.
Summary of ROI -onsiderations
There are many factors in-ol-ed in calculating the *O( of personal data marts. 9rom my perspecti-e! the key factors are as follows) $% Poor cu&e performance is not accepta&le% (f OLAP cubes are needed by an organi5ation! those OLAP cubes need to ha-e excellent performance. )% !here are some inexpensi'e ways to impro'e the performance of #nalysis Ser'er cu&es% Bou should always try these methods to see if they will gi-e the performance you need. *% In some situations( personal data marts and local cu&es pro'ide the fastest possi&le cu&e &rowsing speed% There are expenses in setting up a system of creating and distributing local cube files! but the results are often well worth the expense. 6hen you use personal data marts! you can gi-e indi-idual users a set of cubes that are optimi5ed for the ,uerying they want to do.
-ast 'ubes . appy +sers . More %usiness #nsight
The goal of using personal data marts is to make OLAP fast effecti-e. con-enient! easy! and
9ast cubes make for happy users. (f there are happy users! there will be more users. And with fast cubes! each of those users will be able to find more insights to impro-e the organi5ation.
About the Author
Tim Peterson is the lead author of Microsoft OLAP Unleashed !AM!, "###$ and the author of Microsoft !%L !erver &''' (ata Transformation !ervices !AM!, &'''$) *e
is a spea+er, teacher, and consultant) *e is the chief soft,are architect of -u.e!lice, a tool that automates the creation of local cu.e files)
5ould you li6e to see if the use of local cu&e files in personal data marts could speed up your cu&e &rowsing8 Sign up for a free demo of -u&eSlice now% 0sing remote des6top software( we will show you how to design high performance( customi/ed local cu&e files% You will see for yourself whether they would &e useful in your organi/ation%
-or More #nformation www.cubeslice.com
*o, to (ramatically Improve /ro,sin0 !peed Performance for Microsoft !%L !erver &''1 Analysis !ervices Local -u.es in !%L !erver &''1 Analysis !ervices The Local <ube (nformation <enter
biperformance.spaces.li-e.com
Tim Peterson7s 4( Performance 4log <ube 4rowsing 2peed Performance Local <ube and Other 2trategies to 4oost OLAP
www.olapreport.com
What is OLAP? Performance Matters 8xtensi-e and detailed re-iews of OLAP products
www.microsoft.com
Microsoft !%L !erver &''1 Analysis !ervices Performance 2uide by 8li5abeth Citt. /0uery performance directly impacts the quality of the end user e1perience. As such, it is the primary benchmar2 used to e!aluate the success of an )*AP implementation.3
www.xlcubed.com
/uild /etter -u.es3 4eal5Life Advice on /uildin0 Analysis !ervices <ubes by ?abhan 4erry /4!ery business intelligence solution will ha!e its problems with data scalability. #t"s ine!itable. Almost always, the source data will increase in si5e o!er time. What you ha!e today may perform adequately but this may not be true ne1t month or ne1t year. As technology and hardware has impro!ed, this problem has been alle!iated but not sol!ed6 the limits ha!e simply been mo!ed, not e1tinguished. %uilding a cube that uses all data for all time, and where the data increases o!er time, is a recipe for a cube that will e!entually be too slow to use.3
www.s,lser-eranalysisser-ices.com
*ichard Tkachuk7s Analysis 2er-ices Page! with an article about optimi5ing calculations M(6 Performance *ints
www.ssas+info.com
8-erything about 12 23L 2er-er Analysis 2er-ices %&&. #2uggestion 2elect the menu item !!A! Articles .y !u.7ect and then choose Performance$
www.s,lDunkies.com;6ebLog;mosha
1icrosoft OLAP by 1osha Pasumansky
blogs.msdn.com;s,lcat
1icrosoft 23L 2er-er Ee-elopment <ustomer Ad-isory Team
cwebbbi.spaces.li-e.com
<hris 6ebb7s 4( 4log 1EFtreme Programming
!im Peterson How fast is fast enough for OLAP browsing? OLAP is at its finest when the data changes in less than a second. The user clicks and the data changes instantaneously. This is the kind of performance that allows a user to really explore an OLAP cube drilling up! drilling down! trying one filter! adding an additional filter! switching the dimension displayed on the columns! etc. "igel Pendse wrote an article called Performance Matters for !he O"#P Report #www.olapreport.com$ in August! %&&'. (n it he says the following about OLAP tools) Slow query performance has been consistently the most serious product-related reported problem, and for the last few years it has been the single most often complained of problem. According to data gathered by the OLAP *eport! industry+wide median OLAP ,uery time a-eraged between ' and '.. seconds in %&&/.
Median input data volume and median query performance trends 0sed with permission of The OLAP *eport
1icrosoft 23L 2er-er %&&. Analysis 2er-ices pro-ides excellent OLAP browsing speed. 4ut most organi5ations using Analysis 2er-ices could benefit in certain situations with the performance boost pro-ided by personal data marts. (f your OLAP ,uery response time is o-er sixty seconds! you may be able to bring it down to fi-e seconds. (f you now ha-e a response time of ten seconds! you may be able to reduce it to one second. 6ith faster speed your OLAP cubes will pro-ide more actionable business intelligence because your business users will be willing to examine the data from more perspecti-es.
What Are Personal Data Marts and Why a!en"t # %efore&
eard About $hem
A personal data mart is a set of one or more customi5ed local cube files that ha-e been created for a specific user7s ,ueries. 8ach local cube contains a small portion of the data a-ailable in the Analysis 2er-er cube. 4ecause they are small! they can be -ery fast. 9or example! a company could ha-e a large 2ales cube! with data for :!&&& 2ales *epresentati-es for the past . years. The 2ales *epresentati-es examine the 2ales cube e-ery day. They spend most of their time analy5ing their own sales for the most recent month. A local cube containing only the data for one 2ales *epresentati-e for one month will ha-e only :;/&!&&&th of the data in the Analysis 2er-er cube. 0nused members from the Product dimension! the <ustomer dimension! and other dimensions can be remo-ed from the local cube. This small cube will browse much more ,uickly
than the Analysis 2er-er cube! especially when calculations are being e-aluated. This fast speed can be achie-ed while still keeping all of the cube7s dimensions! attributes! measures! and calculated members. 4esides impro-ing browsing speed! the local cube files also pro-ide the flexibility for the 2ales *epresentati-e to look at data when not connected to the corporate network. Local cube files ha-e been a part of the 1icrosoft 4( toolset since :==> with the release of 23L 2er-er '. Their purpose has been to pro-ide users access to OLAP data while they are offline. The use of local cubes to impro-e cube browsing speed has often not been recogni5ed. 6hile 1icrosoft has always pro-ided the ability to create local cube files! it has often been difficult to find tools to automate the repeated creation of customi5ed local cubes. 6ith the release of 23L 2er-er %&&.! the ability to design local cube files has been impro-ed. The Analysis 2er-ices 2cripting Language #A22L$ gi-es the user full control o-er the creation of the local cube. 1ost users! howe-er! do not ha-e access to a tool that creates local cubes using A22L. Local cube creation tools usually use the <reate ?lobal <ube statement! which gi-es far fewer local cube creation options. A cube created with <reate ?lobal <ube cannot be optimi5ed for browsing speed in the same way as a cube created with A22L. The result is that there are organi5ations with a lot of experience in using 1icrosoft 4usiness (ntelligence! but little experience in using local cube files. 1any of these organi5ations could achie-e a dramatic impro-ement in their cube browsing speed by switching to personal data marts for their OLAP data deli-ery or by using personal data marts for at least some of their more challenging cube browsing situations.
ow 'an # (now #f My )*AP Performance Will #mpro!e With Personal Data Marts&
The only way to know for sure! of course! is to try it. Howe-er! we ha-e found that personal data marts can be particularly useful in the following situations. $% !he cu&es ha'e one or more large( flat dimensions . 2ometimes dimensions are designed without hierarchies or with hierarchies ha-ing -ery few le-els. One example of this is the 2ales Order dimension in the Ad-enture 6orks sample database that is distributed with Analysis 2er-ices. The Analysis 2er-er cannot build effecti-e aggregations for a large! flat dimension. (f you design a local cube that remo-es a significant portion of a large! flat dimension #such as limiting the 2ales Order dimension to the sales for a single 2ales *epresentati-e$! you will greatly impro-e browsing speed when the user is looking at that dimension.
)% !he cu&es use complex calculated mem&ers% <alculations can slow browsing speed. 4ecause there is less data in a local cube! the calculations can be e-aluated more ,uickly. The impro-ed performance is most e-ident in a calculation that has to e-aluate a large number of null cells. 4ecause local cubes can ha-e all unused dimension members remo-ed! the number of null cells can be exponentially reduced. *% Each indi'idual user is using only a small portion of the cu&e . Personal data marts are especially effecti-e at impro-ing browsing speed when each user typically looks at a small subset of the cube7s data like the pre-ious example of a 2ales *epresentati-e looking at the most recent month7s data. +% Some users only want to see an odd su&set of the cu&e% ( ha-e often seen situations where users want to see an oddly shaped subset of the data such as all sales from *egion :! together with the sales from two cities in *egion %! excluding sales initiated by the corporate head,uarters staff! and only including sales to customers who ha-en7t purchased anything in the pre-ious two years. A good OLAP client tool will allow a user to see this subset of the data! but it is much easier and there can be much better performance if the user starts browsing with this exact set of data. ,% -u&e &rowsing speed is ade.uate for some &ut not all users( or is ade.uate at certain times &ut not other times% Analysis 2er-er cubes can be effecti-ely optimi5ed for the typical user! but there may be certain users who are browsing the cubes in unusual ways that greatly increase ,uery response time. 6hen one user issues a long ,uery! the ,uery response time can be slowed for e-eryone using that Analysis 2er-er. (n these situations! it can be -ery helpful to create personal data marts for the users that ha-e the lengthy ,ueries. 2peciali5ed local cubes can be designed to impro-e the browsing speed for these users. 4ecause the Analysis 2er-er no longer has to e-aluate these complex ,ueries! the browsing speed for all the other users can be impro-ed and will be more consistent.
'an # Achie!e the Same *e!el of Performance Without +sing Personal Data Marts&
(n some situations! you can match the performance of personal data marts by using other strategies. (n many cases! howe-er! these other strategies are inade,uate or may be more expensi-e to implement. $% #dd MO"#P aggregations% The Analysis 2er-er gains a lot of its ability to ,uickly respond to ,ueries through the use of aggregations. (f you ha-en7t done so already! create 1OLAP aggregations for your cubes. Aggregations will not! howe-er! increase the browsing speed in large! flat dimensions. They may or may not be helpful in impro-ing the speed of calculated members.
)% Set proper attri&ute relationships in hierarchies . 6hen using the Analysis 2er-er from 23L 2er-er %&&.! aggregations cannot be created for hierarchies unless relationships are set for the attributes used to create the le-els. *% #dd partitions to the cu&e% 6hen used with appropriate slicing in the ,ueries! partitioning can greatly impro-e cube browsing speed. (f there is one single dimension that is causing browsing speed problems! adding partitions along that dimension can be -ery effecti-e. Personal data marts! though! are more effecti-e when different users need a cube optimi5ed along different dimensions. (t7s -ery difficult to take e-ery indi-idual7s situation into account when designing partitions for an Analysis 2er-er cube. +% Optimi/e calculated mem&ers% There are many strategies for optimi5ing calculated members which! in some cases! can dramatically impro-e browsing speed. ,% 0se more powerful hardware% The Analysis 2er-er performs better with a /@+bit ser-er! multiple processors! and a large amount of memory. 1% !each users how to a'oid pro&lem &rowsing areas . 1any Analysis 2er-er cubes will perform well as long as the user doesn7t pick the wrong combinations of attributes. (n many! but not all! situations! slow performance can be a-oided by picking appropriate slicers before attributes are placed on the columns and rows. 2% Simplify the #nalysis Ser'er cu&es. Look at the list of situations where personal data marts can impro-e performance and modify the Analysis 2er-er cube so that these conditions are remo-ed) • • • • • 8liminate large dimensions *emo-e the lowest le-el of detail in large dimensions Add le-els in large! flat dimensions *emo-e calculated members if they are slowing cube browsing *emo-e seldom used attributes
(n summary! all these strategies can speed OLAP browsing. 2ome of them will make a big difference with -ery little effort. 4ut in many cases! the easiest and cheapest option is to gi-e your users personal data marts. There are many excellent resources a-ailable for optimi5ing Analysis 2er-er cubes. 2ee the end of this article for suggestions on papers! web sites! and blogs.
Personal Data Mart ,)#
(n considering the *O( of a personal data mart! there are three primary courses of action) :. 4e satisfied with cubes that ha-e slow browsing speed %. 0se one or more of the other strategies to impro-e cube browsing speed
A. 0se personal data marts to impro-e cube browsing speed 8ach of these options has a cost associated with it.
-ourse of #ction 3$ 4 Be Satisfied 5hen loo6ing at the costs of slow &rowsing speed( consider the following7
$% 5hat is the o'erall 'alue of your Microsoft Business Intelligence system8 How much did your 4( system cost? 6hat benefit is your organi5ation recei-ing from the cubes? 6hat additional benefits are a-ailable through more effecti-e use of the cubes? )% Your cu&es will not gi'e a good ROI unless they wor6 well for your users% 6hate-er benefit you recei-e from the cubes will be greatly enhanced if they can be browsed ,uickly. 9ast browsing cubes are an essential part of a 4usiness (ntelligence system. (t is unwise to spend a lot of money building a 4( system without also achie-ing the best possible cube browsing speed. *% 5hat speed is too slow8 "igel Pendse! in an article entitled What Is OLAP?, has said that a defining characteristic of OLAP is fast browsing response time where simple ,ueries take one second! most responses are returned in about . seconds! and -ery few responses take more than %& seconds. ( think this should be considered to be a minimum standard. 0sers #and organi5ations$ will gain much more benefit from OLAP if the typical ,uery is under : second and -ery few ,ueries exceed . seconds. 6ith this le-el of performance! users feel that they7re getting their ,uestions answered immediately.
-ourse of #ction 3) 4 Other Strategies
The cost is ,uite -aried for the other performance+enhancing strategies) $% Some strategies are 'ery cost9effecti'e% (f you can achie-e excellent cube browsing speed by adding aggregations! setting attribute relations! adding partitions! or by optimi5ing calculated members these would usually be the best options. 8-en if you ha-e to hire a consultant to help you do this! the benefit should be worth the cost. )% Some strategies reduce the 'alue of the cu&es% (f you ha-e to remo-e dimensions! attributes! le-els! or calculations you need to consider the cost of gi-ing up some of the cube7s functionality. Adding le-els to a hierarchy can also make a cube harder to use. The option of teaching users to a-oid certain types of browsing has a -ery high cost because it restricts OLAP browsing to the people who ha-e learned how to do it! rather than ha-ing the benefits of OLAP insight a-ailable to e-eryone in the organi5ation. *% Some strategies can &e expensi'e% The impro-ed hardware option has a definite cost attached to it. And though better hardware can be -ery helpful! by itself hardware may not bring optimal performance.
-ourse of #ction 3* 4 Personal
ata Marts
6hen using the personal data mart strategy! you need to calculate the costs of doing these tasks) $% esigning and creating the local cu&es% (t is fairly easy to create local cube files in 23L 2er-er %&&&. (n 23L 2er-er %&&.! it is usually necessary to use the Analysis 2er-ices 2cripting Language #A22L$ to create optimi5ed local cubes. The A22L can be ,uite difficult to manipulate. Our company has de-eloped a product! <ube2lice! which uses A22L to create customi5ed local cube files that can be optimi5ed for browsing performance. )% istri&uting the local cu&e files. Local cube files need to be recreated when new data is a-ailable. These new local cube files need to be distributed to the users on a regular basis.
Summary of ROI -onsiderations
There are many factors in-ol-ed in calculating the *O( of personal data marts. 9rom my perspecti-e! the key factors are as follows) $% Poor cu&e performance is not accepta&le% (f OLAP cubes are needed by an organi5ation! those OLAP cubes need to ha-e excellent performance. )% !here are some inexpensi'e ways to impro'e the performance of #nalysis Ser'er cu&es% Bou should always try these methods to see if they will gi-e the performance you need. *% In some situations( personal data marts and local cu&es pro'ide the fastest possi&le cu&e &rowsing speed% There are expenses in setting up a system of creating and distributing local cube files! but the results are often well worth the expense. 6hen you use personal data marts! you can gi-e indi-idual users a set of cubes that are optimi5ed for the ,uerying they want to do.
-ast 'ubes . appy +sers . More %usiness #nsight
The goal of using personal data marts is to make OLAP fast effecti-e. con-enient! easy! and
9ast cubes make for happy users. (f there are happy users! there will be more users. And with fast cubes! each of those users will be able to find more insights to impro-e the organi5ation.
About the Author
Tim Peterson is the lead author of Microsoft OLAP Unleashed !AM!, "###$ and the author of Microsoft !%L !erver &''' (ata Transformation !ervices !AM!, &'''$) *e
is a spea+er, teacher, and consultant) *e is the chief soft,are architect of -u.e!lice, a tool that automates the creation of local cu.e files)
5ould you li6e to see if the use of local cu&e files in personal data marts could speed up your cu&e &rowsing8 Sign up for a free demo of -u&eSlice now% 0sing remote des6top software( we will show you how to design high performance( customi/ed local cu&e files% You will see for yourself whether they would &e useful in your organi/ation%
-or More #nformation www.cubeslice.com
*o, to (ramatically Improve /ro,sin0 !peed Performance for Microsoft !%L !erver &''1 Analysis !ervices Local -u.es in !%L !erver &''1 Analysis !ervices The Local <ube (nformation <enter
biperformance.spaces.li-e.com
Tim Peterson7s 4( Performance 4log <ube 4rowsing 2peed Performance Local <ube and Other 2trategies to 4oost OLAP
www.olapreport.com
What is OLAP? Performance Matters 8xtensi-e and detailed re-iews of OLAP products
www.microsoft.com
Microsoft !%L !erver &''1 Analysis !ervices Performance 2uide by 8li5abeth Citt. /0uery performance directly impacts the quality of the end user e1perience. As such, it is the primary benchmar2 used to e!aluate the success of an )*AP implementation.3
www.xlcubed.com
/uild /etter -u.es3 4eal5Life Advice on /uildin0 Analysis !ervices <ubes by ?abhan 4erry /4!ery business intelligence solution will ha!e its problems with data scalability. #t"s ine!itable. Almost always, the source data will increase in si5e o!er time. What you ha!e today may perform adequately but this may not be true ne1t month or ne1t year. As technology and hardware has impro!ed, this problem has been alle!iated but not sol!ed6 the limits ha!e simply been mo!ed, not e1tinguished. %uilding a cube that uses all data for all time, and where the data increases o!er time, is a recipe for a cube that will e!entually be too slow to use.3
www.s,lser-eranalysisser-ices.com
*ichard Tkachuk7s Analysis 2er-ices Page! with an article about optimi5ing calculations M(6 Performance *ints
www.ssas+info.com
8-erything about 12 23L 2er-er Analysis 2er-ices %&&. #2uggestion 2elect the menu item !!A! Articles .y !u.7ect and then choose Performance$
www.s,lDunkies.com;6ebLog;mosha
1icrosoft OLAP by 1osha Pasumansky
blogs.msdn.com;s,lcat
1icrosoft 23L 2er-er Ee-elopment <ustomer Ad-isory Team
cwebbbi.spaces.li-e.com
<hris 6ebb7s 4( 4log 1EFtreme Programming
