-
1. Re: Bad performace in cluster.What is wrong?
michal_szymanski Sep 25, 2010 11:33 AM (in response to michal_szymanski)maybe it is normal write performace, and I expect to much Read performance is very good about 400 thousands get per second on each node.
-
2. Re: Bad performace in cluster.What is wrong?
michal_szymanski Sep 26, 2010 10:22 AM (in response to michal_szymanski)I've found following benchmark for Infinispan http://infinispan.blogspot.com/2010/02/benchmarking-infinispan-and-other-data.html and for 2 node author go 3700 puts ops/sec it is higher than in our test but result is still disappointing (at least for me). It is 20 times less than during this test http://2.bp.blogspot.com/_ca0W9t-Ryos/S4OzJ7jpzQI/AAAAAAAAA_I/yk4-niZmJq0/s1600-h/repl_sync_PUT.png
-
3. Re: Bad performace in cluster.What is wrong?
vblagojevic Sep 27, 2010 2:15 PM (in response to michal_szymanski)You can take our cache bench framework and run on your deployment target. Capture results and report back. Having everyone making their custom performance tests leads to inconclusive results and therefore conclusions.
Regards,
Vladimir
P.S You really node at least several server machines in order to do any decent testing.
-
4. Re: Bad performace in cluster.What is wrong?
michal_szymanski Sep 27, 2010 5:21 PM (in response to vblagojevic)Are you talking about http://infinispan.blogspot.com/2010/02/benchmarking-infinispan-and-other-data.html ? If yes for 2 nodes people get 3700 op/sec what is not huge value. I will try to start it on our server.
-
5. Re: Bad performace in cluster.What is wrong?
vblagojevic Sep 27, 2010 5:47 PM (in response to michal_szymanski)Yes, try it out on 8 servers, or whatever your planned deployment configuration is. That way you will be able to calculate limits of your setup and will not what performance to expect once your application is ready.
Let us know.
-
6. Re: Bad performace in cluster.What is wrong?
michal_szymanski Sep 28, 2010 7:09 AM (in response to vblagojevic)so far we will deploy infinispan on 2 nodes it will be use as cache and replication solution for object that store session information (like credential, information about user it will reduce number of request to DB) for our SOAP WebServices. We will use load balancing and it is a reason why we want to synchronize object in Infinispan between two nodes.
-
7. Re: Bad performace in cluster.What is wrong?
michal_szymanski Oct 1, 2010 3:56 AM (in response to vblagojevic)I'm still waiting for configurationour internal network to perform performacne test using benchmark tool. In meantime I've looked at http://www.jgroups.org/perf/perf2008/Report.html and I still do no understand why in Infinispan on one node performance is about 1000msg/sec but using JGroup it is about 60000 msg/sec 1KB even when I use 4 node (it is our test) I understand that JGroup is only mechanism for communication but it is hard to belive that infinispan waste so much time in inserting data to Cache.
-
8. Re: Bad performace in cluster.What is wrong?
galder.zamarreno Oct 14, 2010 11:12 AM (in response to michal_szymanski)Are you using the same JGroups configuration for both? It's hard to say right now, but if you hook a profiler and run your tests, you'd see where time goes. No speculation beats the result of a profiler . I've used JProfiler in the past and it's well worth the money although there's others out there.
In your environment, you should look at implementing the following recommendations: http://community.jboss.org/docs/DOC-11595
-
9. Re: Bad performace in cluster.What is wrong?
michal_szymanski Oct 14, 2010 1:27 PM (in response to galder.zamarreno)Hi,
there is default configuration, I've described our short test in http://community.jboss.org/thread/157354 you can test in your environment because it is very short program. What do you mean by good performance? I've spent a lot of time to find Infinispan benchmark results but I've found only result on Infinispan blog and results were twice better than ours what is not big difference.
-
10. Re: Bad performace in cluster.What is wrong?
michal_szymanski Oct 15, 2010 4:22 AM (in response to galder.zamarreno)I've run test on two Windows machines and we get 1000-1500put/s in this test CPU load was around 10%. The same test I've started on JProfiler (in this scenario load was 70-80% but it was generated by jprofiler). In attachment you can see results.
Java source of my test:
package pl.carrierex.clusterTest;
import java.io.*;
import java.util.Scanner;
import org.infinispan.Cache;
import org.infinispan.config.*;
import org.infinispan.manager.*;
import org.infinispan.util.concurrent.IsolationLevel;
class Cluster {
public static void main(String[] args) throws IOException {
GlobalConfiguration gc = GlobalConfiguration.getClusteredDefault();
gc.setClusterName("test-cluster1");
Configuration c = new Configuration();
c.setCacheMode(Configuration.CacheMode.REPL_SYNC);
EmbeddedCacheManager cm = new DefaultCacheManager(gc, c);
Cache defaultCache = cm.getCache();
System.out.print("Ver 1.1 t - put 100000 objects\nn -number of keys in memory\n");
String input;
Scanner in = new Scanner(System.in);
while (!(input = in.nextLine()).equals("q") )
{
if(input.equals("n")){
System.out.println("No of keys: "+defaultCache.keySet().size());
} else if(input.equals("t")){
long start = System.currentTimeMillis();
for (int i=0; i<100000; i++)
defaultCache.put(i, i);
System.out.println("Performance= "+((double)100000)
/(((double)(System.currentTimeMillis()-start))/1000)+"put/sec.");
}
}
System.exit(0);
}
} -
11. Re: Bad performace in cluster.What is wrong?
michal_szymanski Oct 26, 2010 9:25 AM (in response to michal_szymanski)I've lost my hope to find a cause of poor performance and we switched to HotRod. By the way with many caches in cluster (and manymodification) we had many error with locks.