atom feed4 messages in org.apache.jakarta.jmeter-userRE: JMeter client performance - a que...
FromSent OnAttachments
André Derrick BalsaJun 5, 2007 6:30 am 
sebbJun 5, 2007 6:39 am 
iblavinsJun 5, 2007 7:13 am 
Sonam ChauhanNov 14, 2007 4:18 pm 
Subject:RE: JMeter client performance - a question
From:Sonam Chauhan (sona@ce.com.au)
Date:Nov 14, 2007 4:18:24 pm
List:org.apache.jakarta.jmeter-user

Just adding to what Ian said....

The key issue is how many maximum _concurrent_ (not total) connections you have.
In my experience things get hairy when handing more than 1000 concurrent threads
in a single Java VM (especially with Java 1.5 due to it's large stack space
allocation per thread). So you may want to startup separate JMeter instances (on
the same box), and spread the load among them. You will probably also have to
tune the JVM heap memory parameters downward in bin/jmeter (or jmeter.bat). That
depends on the complexity of your test.

I've run more than 10,000 concurrent JMeter threads on a single box (Linux on
Xeon with 4GB RAM IIRC), but more than 10 separate JMeter processes were used to
drive that load.

-----Original Message----- From: iblavins [mailto:ibla@temenos.com] Sent: Wednesday, 6 June 2007 12:14 AM To: 'JMeter Users List' Subject: RE: JMeter client performance - a question

G'day

JMeter is inherently capable of running 6,000 sessions through a server from a single instance of JMeter.

Whether it will or not in your circumstances depends on a number of things.

Firstly how many of the 6000 connections will be concurrent. This is dependent on how long, on average, they will run and the rate at which they start. So if all start together and run for 5 minutes then concurrency will be 6000. If you start 1500 per minute and they run for 1 minute then the concurrency will be around 1500. The higher the concurrency the larger the JMeter box required to run the workload.

Secondly it depends on the rate that each session makes calls to the server ie how active it is. The more server calls per session per minute the bigger the JMeter box needs to be.

Thirdly it depends on how heavy your samplers are. I have seen JMeter run over a million samplers per minute from a single machine when the samplers do nothing. But add in some complexity in the test plan, a couple of remote server calls to get data for the sampler and to log the results of the sampler, and the throughput is more in the 50,000 samples per minute on the same hardware.

Fourthly it depends on the size of the LINUX box. The above results were under AIX on a 3 * 1.65GHz twin core CPU LPAR of an IBM p570 server. This is a pretty strong box - it has a rating of something like 108 on the spec_intRate2000 scale at www.spec.org.

But, broadly, at the 6,000 session mark, assuming concurrency is 2,000 or more, I'd be assuming you will need multiple JMeter servers unless your LINUX box is a big one.

Ian Blavins Contract Performance Engineer Temenos

-----Original Message----- From: André Derrick Balsa [mailto:andr@gmail.com] Sent: 05 June 2007 14:31 To: jmet@jakarta.apache.org Subject: JMeter client performance - a question

Hello,

I am in charge of load testing a Weblogic/Oracle system with quite a few servers and a relatively complex application.

The requirement is that the system should be able to handle 6000 (six thousand) connections over a five-minutes period.

My question is : can a single JMeter client running on Linux handle 6000 connections over a five minutes period ? Or do I have to distribute this over a few machines ?

Can anybody who has had some experience running JMeter on a Linux client machine share her/his experience ?

Thanks a bunch, Andrew