If you're configuring DataDirect ODBC drivers on *nix system, be aware of following:

1. There are 2 versions of odbc.ini file, one in 32bit driver folder, other in 64. 64 essbase uses 64 drivers.
2. You should copy required odbc.ini file to essbasepath\bin. And you should create a symbolic link (ln -s) to odbc.ini file in essbase user home folder (/home/hyperion, for example)
3. When creating a SQL data source rule file -- ServerName corresponds to database server name, not essbase server name (type TNS name for oracle database)

Useful links:

Essbase SQL Interface Documentation

Metalink on search keywords 'odbc.ini + essbase' )

Oracle BI EE Pro's:

1) Aggregate navigation

-- ability to set up aggregate tables in BI EE itself, which gives it 'aggregate awareness' while generating SQL. So if you have monthly sales aggregate you just set up BI metadata accordingly and all reports will target this table for monthly data.
There's no such feature in Cognos metadata setting.
But using this feature arises some questions:
a) whether aggregate definition should be BI-tool specific. A lot of practioners, including Kimball, insist that aggregate navigator should reside at database level so that every tool querying datawarehouse will benefit from aggregate avalaibility. This seems rather reasonable, since every datawarehouse is usually queried by more than one analytical application (BI, datamining, more BI) so you either direct all applications to BI EE, or agree to performance degradation.
b) aggregate tables desynch. When you just update the base fact table, aggregate tables become "out of synch", providing incorrect query results untill they're recalculated. This means that you either can guarantee that nobody will access reports in this period, or you can have incorrect data. Since more&more datawarehouses squezee load windows to reach real-time this problem gains priority.
c) choosing which agregate to use for answering questions. That's what statistics is all about in dbms, knowing number of rows and value distribution (lol). Oracle BI EE has the "table row count" feature (although I haven't seen it affect SQL generation yet, need more examples). But there's no value distribution analysis in there, so it's just one side of the coin (not talking about I\O device speed and other characteristics).

In general, it's recommended to use database-specific aggregate table functionality (Oracle Materialized Views or DB2 MQT) since it solves all the 3 questions given above and usually simplifies ETL process (and sometimes even speeds it up, since databases use their own transaction logs to detect what data has changed and what aggregates should be rebuilt). Too bad indexed views in Ms SQL do not work with aggregate dimensions (there's no way to define dimension hierachies there).

2) Cache management

-- OraBI has really profound cache management facility. You can "cache" any database query to OraBI specific storage structure (cache file), wich will allow subsequent queries to the same table\query to run without actual database request being made. This can greatly speed things up. I especially like the Event Polling Table feature: you add a table, which records when dwh table was last updated. OraBI then reads at given intervals and automatically invalidates old cache entries, based on this table records.
Moreover, if you have OraBI cluster this cache can be shared among servers.

There isn't anything even close in Cognos.

Although I greatly like this feature, I just want to warn about overusing it. It's easy to imagine BI developers boosting performance by adding more&more cache untill OraBI becomes a fully blown aggregate system. And sometimes it's just about 1 aggregate table at the dwh level ) Or about introducing OLAP server in the enviroment )))

3) SQL generation.

It's tricky subject, but for now I like OraBI generated SQL more. But it's "apples to oranges" for sure, since I usually use DMR's in Cognos which encumbers SQL greatly and there's nothing identical in OraBI EE.

Oracle BI EE Con's:

1) Multidimensionality

Cognos has Analysis studio and the ability to navigate hierarchies in both directions (you won't believe it, but in OraBI there's no 'Drill-Up', only'Drill-Down'). And DMR's and analytical functions (but their usage is a bit annoying, as it seems now, hope to write about it later). Anyway, Cognos is way much more 'multidimensionally-ready' than OraBI.

OraBI + Essbase is a work in progress and has a huge number of caveats (some fixed by patches, some not, some introduced). And the only way to use all Essbase functionality is to write direct MDX via Evaluate functions. That's a big problem, since it's hard for us to suggest OraBI on top of Essbase for now (till 11g once again). The only alternative is Visual Explorer, which is a very good tool, but it for top-analysts only (thick client, costly).

2) Metadata model development

Instead of OraBI's 'only-star schema', '3 layers of model' Cognos FM Manager doesn't impose any design principles, which allows more mistakes, but it makes some things way more simple. One of the first things I wanted to do in OraBI was a report using just a single table.  Well that's a really funny exercise (see posts over here) if a star schema is a must.

But the main problem is the lack of API for metadata changing and browsing. There's udml, but it's not supported officially. Therefore all current integration scripts (like adding users, merging repositories and working with hierarchy depth changes) are out of the law. Which doesn't stop anyone, but is pretty annoying.

3) Pixel-perfect reports

Oracle BI Publisher is a specific tool, aimed at generating a huge number of formatted reports, based on XML format definition files. It wasn't a part of Siebel BI, so 'integration stitches' still stand out. It's a nice tool, but it certainly lacks web-interface ) Therefore in Oracle BI there's a deep distinction between a simple formatted report (with lots of possible logic in it) and making this report 'printer-friendly' since for the latter you basically have to start from scratch by opening Ms Word )  This will change in 11g as they say )

Having said all that, I really wait for Oracle BI 11g edition to start using it with Essbase and I kinda like the product as it is for "relational-only" reporting.

I surely wanted to write a simple bullet point list at first )

The main description of ASO storage is "data is stored as key\value pairs". This quote comes from ASO Tuning WhitePaper, the only "internals" technical description of ASO out there. It's for 7, but all the concepts are still valid.

So, "key-value pairs" it is. Key length is very important then, because it directly affects:

a) size of the database, if you've got a billion rows cube, then -1 byte for every cell means 1 Gb less overall cube size (that's raw size, it'll be compressed later). 
b) query processing -- for every query cell key got to be calculated for data retrieval. Less key, faster the query, as I think.

You can see the key description on Application Properties page or issuing "query database test.db list aggregate_storage runtime_info;"  MaxL command .
Here's the sample output. 

parameter                                         value                                            
+-------------------------------------------------+-------------------------------------------------
 Dimension [Date] has [3] levels, bits used                                                        1
 Max. key length (bits)                                                                            1
 Max. key length (bytes)                                                                           8


we see following characteristics of dimension -- number of levels and number of bits used to encode values in this dimension. Cell key is concantenation of each dimension keys. 

But that number of bits is pretty cryptic, though. On a project we're doing we have a couple of 5 mln elements dimensions and key length for them is around 40 bits. If you use simple binary encoding -- you can encode 2^40 elements (that's really a huge number, it contains 14 digits). It's no a simple binary encoding then. 

So how are dimension elements encoded? 

Well, it's allways easy afterwards, so I'm not so proud now, as I was when it first struck me )

So I now think that the technique called hierarchial encoding (or indexing) or something similar is used, so each dimension element is encoded the following way:

Let's assume that we have 3 levels in dimension, then the key will be formed like:
xxxyyzzzz
where xs -- are the bits requiered to binary encode level 2 parent
ys -- level 1 parent
zs -- the bottom level element

So to form the key length for all dimension you have to concantenate binary keys required to encode elements on every level. 

Therefore there are a few things to keep in mind while doing big-scale ASO projects:
- think about number of levels in big dimensions, keeping them short greatly decreases size
- try to keep number of elements even on every level -- that's funny, but it'll help to fully "pack the key" 

But that brings up a few more things about ASO i'm thinking of:
- how are pages in tablespace filled for data storage? I think some sort of hierarchial clustering, using key parts to point at pages
- what and how is stored in the outline (otl file)? Membernames and aliases for sure, but also the dimension elements keys, which are concantenated to get cell key and find the data page in tablespace.

 
So I'm still waiting for Edward to publish ASO book, there's a lot of questions. The book seems to be close )

Those who've listened to our recent Oracle events presentations (russian TechForum 2008, BI Forum 2009), know that we've made a nice PoC project with SportMaster, consisting of loading their year's worth of daily stock and sales data (around billion fact cells with over a million elements in products dimension).  All in all, I've decided to post a write-up about how to achieve acceptable load speed on such volumes and tell a scary-story, with a detective twist and a happy ending.

I won't go in much detail on basics, so:

Since it's a billion cells cube, it's ASO storage option.

Since it's an ASO Cube, we use load buffers for data loading.

Since we use load buffers, we try to split input data into chunks, to maximize data read speed.

Since we use load buffers and we were doing PoC on 9.3.1, we have to extract data from DWH to text files (in 11 we could use ODBC based load buffers).

Just a remark: Essbase Integration Services does a lousy job on either building dimensions (we're adding 14 attributes for products and loading a file again for _every_ attribute is a disaster) or loading data (no visible ability to run multiple buffer load). No comments on 11 Essbase Studio, haven't tried it.

So, we've extracted data from DWH to text files and are loading these files into text buffers. While loading data from file into buffer, with a simple rule file I could achive data read speed of around 6 Mb\s (on a 80 Mb\s disk system, we've got two such disk systems attached on PoC server). Loading two files in parallel, gave 12 Mb\s. So to fully load up disk system we need to chunk files into 6 parts. That actually doesn't present a problem, but seems a bit awkward. While testing, I've accidentially forgot "using server rules_file" in MaxL script and was astonished by load speed of 25-40 Mb\s! That's how I've found out about 'free-form data load' )

Essbase has a built in ability to load data without rule file (see DBAG). And, moreover, this is almost 10 times faster than loading with rule file. Easiest way to find more about free-form load is to export an ASO cube to text file (right-click on database in Essbase Administration console). Moreover, free-form allows to load data in cross-tab (pivot accounts in columns), therefore greatly suppressing number of rows. Using this technique requires additional data formating: spaces as separator, double quotes enclosure, writing "Missing" or "MI" for null values. But load speed is worth it )

So the bottom line: free-form data load, additional data massage -- 70 Gbs of data in less than 30 minutes, almost reaching half of disk system speed.

That's it for fast data loading, if you're on 9.3.1 i'd recommend trying to massage data into free-form and load it this way.

Now for detective part of the story.

So we've reached ultra-high loading speed and loaded over-half a billion rows in less than half an hour (+ another 40 minutes for aggregation, to be sincere). And then left it out for a couple of months. Fast forward some server migration, another PoC and all. Then once we've need those PoC results once more -- they dissapeared. But that's not a mystery, this usually happens on dev servers. Backups were gone as well, but that's common for dev as well.

But then I've tried to reload data -- it just didn't. Said -- 'unknown member' on a member that was in outline and stopped loading. Okay, let's try to focus on finding error rows. So it's chunking the 25 g file into smaller ones and trying again. 1mln rows -- doesn't work. 100k rows -- no go. 10k -- no go. 1k -- no go. 500 -- YEP. Excellent, so it's about those other 500 rows. Adding another 100 -- NO. 50 -- YEP. 75 -- YEP. 100 -- NO. 90 -- NO. Hm, strange. Another 10 rows -- YEP.  What? I've already loaded 100 and got NO. 100 again -- NO. Again -- YEP. That's when I blew my head off. I've got a data sample that I could load successfully one time out of 4. That's crazy.

Rule file load worked. But ten times as slow.

I'll add some white-lines for those with essbase experience to think about why this error could happen ) Loading the same chunk of data either gives 'unknown member' error or loads successfully.

Looking back it simple ) -- this depends on whether Essbase can find selected member name in outline cache or no. By default all member names are preloaded into memory on application start, but there's an essbase.cfg setting PRELOADMEMBERNAMESPACE which allows to turn off this loading (which was done by someone between PoCs), If preloading is turned off member names are loaded the same way as the whole outline -- by chunks of OUTLINECACHESIZE kbs (by the way, if somebody read to this part -- there's a free beer for you). Off: Anybody knows why outline paging cache size setting officially dissapeared from documentation after 9.2.1??. And free-form loading is so fast, that outline cannot be loaded on same speed to get member names (they can be in different parts of otl file), so it eventually fails with 'unknown member'. Some sequential tries, however, lead to exactly needed piece of otl lying there in outlinecache, so the whole file loads successfully.

The answer was to turn member names prealoading back on -- it started to work again. But tracking the error was really messy.

Lessons learned: Allways check essbase.cfg. And backup everything )