[infinispan-dev] Infinispan Large Scale support

[Bela Ban]

I cross-posted this to the JGroups mailing lists [1]


[1] https://sourceforge.net/mail/?group_id=6081


On 3/22/11 2:05 AM, Dave wrote:
> I switched back to UDP today based on your feedback. Our config resembles
> the config below. Like I said we just increased sizes and timeouts. If you
> ask me why I tweaked a certain parameter my response would be that it seemed
> like a good idea based on the JGroups documentation. UDP seemed a little
> more problematic than TCP, not sure why though.
>
> <config xmlns="urn:org:jgroups"
>          xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
>          xsi:schemaLocation="urn:org:jgroups file:schema/JGroups-2.8.xsd">
>     <UDP
>           mcast_addr="${jgroups.udp.mcast_addr:228.6.7.8}"
>           mcast_port="${jgroups.udp.mcast_port:46655}"
>           tos="8"
>           ucast_recv_buf_size="20000000"
>           ucast_send_buf_size="640000"
>           mcast_recv_buf_size="25000000"
>           mcast_send_buf_size="640000"
>           loopback="true"
>           discard_incompatible_packets="true"
>           max_bundle_size="4000000"
>           max_bundle_timeout="30"
>           ip_ttl="${jgroups.udp.ip_ttl:2}"
>           enable_bundling="true"
>           enable_diagnostics="false"
>
>           thread_naming_pattern="pl"
>
>           thread_pool.enabled="true"
>           thread_pool.min_threads="2"
>           thread_pool.max_threads="30"
>           thread_pool.keep_alive_time="5000"
>           thread_pool.queue_enabled="true"
>           thread_pool.queue_max_size="1000"
>           thread_pool.rejection_policy="Discard"
>
>           oob_thread_pool.enabled="true"
>           oob_thread_pool.min_threads="2"
>           oob_thread_pool.max_threads="30"
>           oob_thread_pool.keep_alive_time="5000"
>           oob_thread_pool.queue_enabled="true"
>           oob_thread_pool.queue_max_size="1000"
>           oob_thread_pool.rejection_policy="Discard"
>           />
>
>     <PING timeout="360000" num_initial_members="400"
> break_on_coord_rsp="false"/>
>     <MERGE2 max_interval="30000" min_interval="10000"/>
>     <FD_SOCK/>
>     <FD_ALL/>
>     <BARRIER />
>     <pbcast.NAKACK use_stats_for_retransmission="false"
>                     exponential_backoff="0"
>                     use_mcast_xmit="true" gc_lag="0"
>                     retransmit_timeout="300,600,1200,2400,3600,4800"
>                     discard_delivered_msgs="true"/>
>     <UNICAST timeout="300,600,1200,2400,3600,4800"/>
>     <pbcast.STABLE stability_delay="1000" desired_avg_gossip="50000"
> max_bytes="1000000"/>
>     <pbcast.GMS print_local_addr="false" join_timeout="60000"
> view_bundling="true" use_flush_if_present="false"/>
>     <UFC max_credits="2000000" min_threshold="0.20"/>
>     <MFC max_credits="2000000" min_threshold="0.20"/>
>     <FRAG2 frag_size="2000000"  />
>     <pbcast.STREAMING_STATE_TRANSFER/>
>     <!--<pbcast.STATE_TRANSFER/>  -->
>     <pbcast.FLUSH timeout="0"/>
> </config>
>
>
> -----Original Message-----
> From: infinispan-dev-bounces at lists.jboss.org
> [mailto:infinispan-dev-bounces at lists.jboss.org] On Behalf Of Bela Ban
> Sent: Saturday, March 19, 2011 1:15 PM
> To: infinispan-dev at lists.jboss.org
> Subject: Re: [infinispan-dev] Infinispan Large Scale support
>
> Hard to believe that TCP would be better, as TCP creates a mesh of
> connections; for 400 nodes, with every node sending, you'll have roughly
> 400*400 connections !
>
> I always had a much better experience with UDP
>
> On 3/19/11 2:37 PM, david marion wrote:
>>
>> Initially yes, but I think we are getting better stability using TCP. I
> switched it back to TCP yesterday. I can post specifics of what I did in the
> TCP configuration, but the short story is I increased a lot of the timeout
> values to get it to work.
>>
>> Dave Marion
>>
>>
>>> Date: Sat, 19 Mar 2011 10:50:54 +0100
>>> From: bban at redhat.com
>>> To: infinispan-dev at lists.jboss.org
>>> Subject: Re: [infinispan-dev] Infinispan Large Scale support
>>>
>>>
>>>
>>> On 3/18/11 10:35 PM, Dave wrote:
>>>> Won't be able to get CR4 uploaded, policy dictates that I wait until
> final
>>>> release. However, I was able to get 431 nodes up and running as a
> replicated
>>>> cluster and 115 nodes up as a distributed cluster. For the 430 node
> cache, I
>>>> was able to get it started with no problems about 50% of the time. When
> they
>>>> formed multiple clusters they merged together only some of the time. It
>>>> really does appear to be a startup issue at this point. We have not
> pushed
>>>> it hard enough yet to see what happens at this scale under load.
>>>>
>>>>
>>>>
>>>> Any idea when CR4 will be FINAL?
>>>>
>>>> Are there any tools to help diagnose problems / performance at this
> scale (I
>>>> ended up writing my own monitor program)?
>>>
>>>
>>> Yes, there's probe.sh at the JGroups level. I created a JIRA to provide
>>> a sample for large clusters. You said you based your config on udp.xml,
>>> correct ?
>>>
>>> [1] https://issues.jboss.org/browse/JGRP-1307
>>>
>>> --
>>> Bela Ban
>>> Lead JGroups / Clustering Team
>>> JBoss
>>> _______________________________________________
>>> infinispan-dev mailing list
>>> infinispan-dev at lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>    		 	   		
>>
>>
>>
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>

-- 
Bela Ban
Lead JGroups / Clustering Team
JBoss