400 caches equals 400 eviction threads ?

[tlvenn]

Hi,

I am starting a new topic for this question alone that I raised in this topic : http://community.jboss.org/thread/157035?tstart=0

Right now, I got close to 400 caches and infinispan spawn one eviction thread per cache....

Sanne suggested to use LIRS instead of LRU to address the issue, totally forgot about that one and I then dig a little bit in the doc and found out also that i should be able to use the new thread-policy attribute of the eviction element to "PIGGYBACK" to avoid this as well.

The issue is I tried both and no matter what, I still end up with the 400 eviction threads. I am using IS-4.2.0-ALPHA2 and HIB-3.5.6

Thanks in advance for your help !

--

Chris

[tlvenn]

FYI, same thing happen with IS-4.2.0-ALPHA3,

All my named cached and the default have the following eviction:

<eviction wakeUpInterval="5000" maxEntries="10000" strategy="LIRS" threadPolicy="PIGGYBACK"/>

Any ideas ?

[tlvenn]

Apologize in advance for bumping this topic but I really would like to solve this as right now all those threads are taking quite some ressources and I am pretty sure that the fix is simple but just happen to evade me for now...

Anyone got an idea of what might happen ? Any leads ?

Thanks !

[vblagojevic]

I believe there is one dedicated thread for eviction per cache no matter what policy you set.  Maybe we should revisit this design decision. I'll have a look.

Regards,

Vladimir

[tlvenn]

Hi Vlad,

Yep looks like that even with piggyback, a dedicated thread is still allocated to handle expired elements out of the given cache. Ideally when the thread policy is set to piggyback, a common pool shared by all caches should be used to remove expired elements instead of a dedicated one per cache.

Thanks for looking into this !

--

Chris

[galder.zamarreno]

Chris, fyi: https://jira.jboss.org/browse/ISPN-730

You can find a link to the infinispan development mailing list where we discussed your use case.

[tlvenn]

Thanks Glader and all the devs who participated to the topic here and on the mailing list. Until the issue is adressed, I will look into creating a custom ScheduledExecutorFactory as suggested by Mircea.