Skip to Main Content

Breadcrumb

Question and Answer

Tom Kyte

Thanks for the question, sean.

Asked: March 31, 2002 - 5:34 pm UTC

Last updated: March 08, 2007 - 3:37 pm UTC

Version: 817

Viewed 1000+ times

You Asked

I have been thinking of the setup of emulating many and many concurrent users. I came across reading Jonathan Lewis'Practical Oracle8i', on the page 4:
"
If you stop to think about it for a minute you will apprecite that these features are pretty astounding, especially when you consider that Oraccle has published details of a bechmark emulating 10,000 concurrent users ( and I have heard mention of one benchmark hitting 50,000 concurrent users).
"

Do you know what the becnmarks he refered to? Or what are your thoughts on the implementation of emulating a lot of concurrent users?

Thanks

and Tom said...

Well, 10k is a good sized but smallish system these days. Our email servers internally handle more then that on a regular basis.

The 50k benchmark is *probably* the one we did when 8i first came out with cache fusion for OPS. We did a stress test with 50k users -- it was part of the 8i rollout a couple of years ago.

Check out:
</code> http://www.tpc.org/results/individual_results/HP/tpcc_hp_sd_750_es.pdf <code>

for a number that might make you think twice ;)

To emulate lots of clients you need some load testing software like loadrunner and the system you want to test... Make sure to beef up the clients!



Rating

  (9 ratings)

Is this answer out of date? If it is, please let us know via a Comment

Comments

any other clients,

sns, January 10, 2006 - 3:09 pm UTC

Tom,

to emulate 10K users login simulteneously and execute certain SQL's, do you know any tool which you are aware of?

I searched in google and found quite a lot which I am not sure if it satisfies my purpose.

Is there a way to emulate 10K users without using any tool?

Thanks

Tom Kyte
January 10, 2006 - 3:14 pm UTC

without using a tool? only by developing your own.

Mercury Interactives Loadrunner is pretty good.

If you have 10,000 simultaneous (concurrently active) users, you have a fairly large system - it would be worth your money to invest in such a tool.

develop on my own,

sns, January 10, 2006 - 3:32 pm UTC

Tom,

thanks for the response. However, it is upto the management to decide about buying the tool.

I am interested to know how to develop on my own. If you have any guidelines or ideas could you please share with me?

Thanks,

Tom Kyte
January 10, 2006 - 4:29 pm UTC

it would take quite a bit of effort.

first you need enough hardware to run 10,000 clients.

then you need the software infrastructure to start up 10,000 client.

you need to reporting system so you get some metrics out of this.

and so on.


google around - you'll find sourceforge like projects out there already existing.

JMeter

TBarry, January 10, 2006 - 4:22 pm UTC

If you can work with Java, JMeter is an open-source tool from the Apache Jakarta project. I don't have a ton of experience with it, but seems pretty flexible. You can load-test web apps and you can load-test database access, but that access will be through JDBC.

</code> http://jakarta.apache.org/jmeter/index.html <code>

emulating users,

sns, January 10, 2006 - 4:54 pm UTC

Well 10K is probably a big number we eventually want to try. But surely we want to try 200-300 users at first and then increasing the number.

So for 200-300 users (clients), is there a simple way?

Thanks,



Tom Kyte
January 10, 2006 - 7:30 pm UTC

take heed from the poster right above? or google about, you'll stumble upon many things.

I just plugged in "load test tool" to google and got what looks like a really good start.

Is it Possible using DBMS_JOB?

A reader, January 10, 2006 - 10:43 pm UTC

Tom,
I was reading one book on PL/SQL and found that similar testing was done using DBMS_JOB.SUBMIT. Is it possible to simulate this requirement with DBMS_JOB?

Thanks.

Tom Kyte
January 12, 2006 - 10:02 am UTC

not realistically.

first, job_queue_processes is 1,000 maximum.

second, everything would take place on the database server - this is unlikely to mimick your actual implementation.

load testing it used to test your system - if you have clients submitting sql from another machine, running everything (the client and all) on the same machine with the database is not realistic.

grinder

anil, January 11, 2006 - 3:50 am UTC

hi .
there is a free tool called Grinder that we have used recently to simulate up to 3000 concurrent users. IT is very good

rgds
anil

Thanks guys,

sns, January 11, 2006 - 10:36 am UTC

One more thing is I don't want my webserver to put into load stress test. I have SQL statements that runs against the database and I want just my database to be put under load stress test by emulating whatever number of users.

Anil: Thanks for letting me know about GRINDER.

Thanks,

Tom Kyte
January 12, 2006 - 10:35 am UTC


A reader, January 12, 2006 - 3:57 am UTC


stress testing

ignorant, March 08, 2007 - 2:48 pm UTC

Hi Tom,

I absolutely subscribe to your view of stress testing the application, not the database/hardware. However in some cases we can only try to get as close to that ideal as possible.

Consider this, I am working on a 9i system that is about to go live. The problem with that is that I can roughly estimate loads, etc based on the current application's load but it is not going to help much. In other words, I need to throw that out the window (except number of concurrent users) and devise a stress test before it goes live.

In the absence of Loadrunner or any such tool (won¿t get money for it), what do you think about the following approach -?
1) Get one user to log on the system.
2) Find the SID and trace it.
3) Get the user to do all regular "expected" tasks. This has to be a small enough subset of what the clients will ultimately do but should model all critical processes.
4) Stop trace.
5) Modify the sql statements to 'parameterize' them and put them into some sort of PL/SQL block.
6) Fire this procedure off with different values (random) up to no_of_concurrent_users (from previous system) + 10% overhead.

Drawbacks:
1) Clearly the time to come up with this will be quite long. I am looking for any recommendations that you might have to reduce the effort (capturing queries into table, etc).
2) This does not accurately model what will happen because the queries remain the same once compiled. This might not happen in the "real" scenario.
3) This approach ignores SQL message times, etc that might be relevant when the production switches over.

Given these drawbacks, it should still help in finding out ugly queries and disk contention. But is it worth the effort?

What is your opinion? Also, can this process be streamlined a little more?

Yours truly,
ignorant.
Tom Kyte
March 08, 2007 - 3:37 pm UTC

testing is not a trivial task - many people never schedule the time for it. It takes a significant amount of effort and (if you ask me) should be part of the development schedule from day one (rarely is). That is, as the guys writing the code are writing the code - they are also thinking about how to test drive this code and building that infrastructure.

Your approach is "perhaps OK, but won't be entirely real world". I'm not sure it will work however - how do you parameterize that working set with inputs that actually "work" (do not result in errors/failure). That'll be tricky.

Sorry, no magical silver bullets here.
more

Connor and Chris don't just spend all day on AskTOM. You can also catch regular content via Connor's blog and Chris's blog. Or if video is more your thing, check out Connor's channel and Chris's channel from their Youtube channels. And of course, keep up to date with AskTOM via the official twitter account.