Conference nsic00::eis_dw

484.0. "100 simultaneous users on a brazilian Dataware ..." by VAXRIO::LEO () Fri Apr 25 1997 16:18

Hi,

	This is Leo from Digital Brazil.

	I am working on a very big Datawarehouse opportunity on SERPRO, the 
brazilian company that processes all the government budget.

	They are intending to upgrade their Digital equipment in order to be
able to support more users.

	Today they have 10 simultaneous users.

	The Digital Equipment is an AlphaServer 4100 using 3 cpu 5/400 and 
3 GB of RAM. They are using HSZ40 with 5 disks in each one of the 6 
channels.

	The Oracle SGA is using 2GB of RAM. 

	They are using DSS client software from MicroStrategy. DSS generates
SQL queries that are executed against a 60 GB database with some 4GB and 8GB
fact tables.

	The queries are typical datawarehouse ones. We are optmizing the queries
using Oracle hints and the database is configured in order to support a
datawarehousing environment.


	The average response time, that each one of the 10 simultaneous users
get is 30 minutes, for each individual querie.

	The CPU is completely busy and the memory becomes very used during
some very heavy sort operations.

	I know that the question I will pose is  very hard to be answered with 
a good precision.

	But the customer needs to have at least an aproximate estimative 

	He wants to be able to support 100 simultaneous users and they are 
absolutely sure that they will need more CPUs, more meory and so on.


	The profile of the queries will still exactly the same but then 100 
of them will be executed in a simultaneous mode.  


	I am strongly considering that I will need a Turbolaser 8400 in order
to support this king of workload.

	But does anyone has any idea in terms of numbers of 5/440 cpus and 
MB of RAM that could be able to mantain the response time of each one of the
100 simultaneous queries in 30 minutes ?

	I think that this environment is a typical Datawarehouse picture and 
any info about hardware configuration used to support similar environment
all over the world would be very appreciated.

	It's very important to know that so far VLM is not being used because
the application  needs a lot of PGA space that not fits completely on the 2GB
SGA. Therefore we need to have some free memory available after starting 
the 2GB SGA. Althought we are using a Multi-Threaded Oracle Server and all the 
session information (part of PGA) is already mapped into the SGA we are still 
needing  to reserve some memory besides the SGA (2GB) to the PGA.

	We also need some external SGA memory in order to improve several 
sort operations.

	If we were using more MB of RAM to SGA (for instance 6GB or 8GB) what 
kind of improvement would be verified in terms of response time ?

	I know that is impossible to have an "black or white" answer to that
but any estimative or information of similar workload environment would help
us very much in order to achieve the best price/performance configuration
on our biggest Datawarehose case, down here Brazil.