Re: 9i server performance, 5 million hits per second?

Nice line of analysis -

I stopped after using my (very simple, rather crude) approximation that you can get

    10,000 logical I/Os per sec per 100MHz of CPU

    5,000,000 queries per sec =

        15,000,000 LI/Os per sec (per 3 blocks) =
            1,500 x 100MHz of CPU =
                150 GHz of CPU

Looks like a 75-way NT box running 2GHz chips, or does anyone do a box with more CPUs ? Maybe the old Pyramid RM1000 with 1,024 x 200MHz chips ?

Then you have to start allowing time for contention

Is it possible the intention was 5M queries per hour ? Unless my arithmetic skills are really packing up, to get 5M per second, every person on the planet would need to fire off a query every 20 minutes ;)

--
Jonathan Lewis
http://www.jlcomp.demon.co.uk

Now running 3-day intensive seminars
http://www.jlcomp.demon.co.uk/seminar.html

Host to The Co-Operative Oracle Users' FAQ
http://www.jlcomp.demon.co.uk/faq/ind_faq.html

Author of:
Practical Oracle 8i: Building Efficient Databases


Andrew Mobbs wrote in message ...


>Just consider memory use. You need to access a minimum of 3 blocks

>(index root, index leaf and data), you need to take out latches on all

>the buffers, then free them; at least read and calculate the hash of the

>SQL string and check it against the library cache, and dozens of other

>bits of housekeeping.

>

>This all adds up to a few kB per query that Oracle needs to transfer

>from main memory to the CPU. Think about a few multiples of "a few kB"

>by your target of 5,000,000 queries per second:

>

> 8kB -  40 GB/s

>16kB -  80 GB/s

>32kB - 160 GB/s

>

>I'm not going to guess where in that range a single query lies.

>

>To pick an example of a current top-end server, Compaq's marketing for

>the GS320 claims "Over 51GB/s aggregate internal bandwidth". Dilute

>that to taste for figures achievable with a real application in the

>field.

>