Re: Load testing for Users

Well, surely you can calculate some things.

How big will the PGA for each user be? Now multiply this by 1000 users. Add the SGA size. Add the other memory requirements. Do you have this much real memory? If not, swapping / paging will occur, so how fast is your swapping / paging device?

Suppose your query is properly indexed -- that is, all conditions are satisfied in the index. So each user may need 1 or 2 I/O's to retrieve the index leaf with his own key. Then each user will need up to x I/O's to retrieve the blocks containing his rows. I say this because there is no reason to believe that any user's x rows are clustered into the same block -- unless you have forced that to be true.

Likewise, there is no reason to believe that a block retrieved for one user will contain a row of interest to another user. Yes, there will be some, but probably few cache hits here!

So we are looking at (x + 2) * 1000 random I/O's. How fast is your disk for random reads? A good average for SCSI disks is 80 random I/O's per second. So I guess you are in the ballpark of (x+2)*1000/80 seconds of total elapsed time. Maybe a little less if some of the index I/O overlaps the data I/O. If x = 6, that is about 100 seconds.

sanjayraj6844_at_my-deja.com wrote:

> I have an Oracle 7.3.2 database running on HP-UX machine. The database
> has approximately 40 base tables. Some of the tables have approximately
> 10 million rows. I wanted to test the effect on performance when 1000
> users simultaneously run queries against a particular table having 2
> million records). The query can be as simple as SELECTING x number of
> records pertaing to the particular user i.e. all the queries will be
> similar excepting say the USERID field. Is it possible to simulate such
> a senario. I know I can use Mercury Interactive's LoadRunner or similar
> tool but I donot have access to any such tool. Can anybody help.
> Pl copy me at sanjay.raj_at_roche.com
>
> Thanks
>
> Sent via Deja.com http://www.deja.com/
> Share what you know. Learn what you don't.
Received on Thu Jun 10 1999 - 10:09:11 CDT