Found the reason for many evictions (bug in our code), please ignore the 
specific question on filter cache.
All other questions (in bold) are still very relevant.



From:   Esther Goldbraich/Haifa/IBM
To:     solr-user@lucene.apache.org
Date:   27/08/2015 01:13 PM
Subject:        Re: Solr 5.2.1 versus Solr 4.7.0 performance


:)

Yes, 64 collections / 1 shard is compared to 1 collection / 64 shards 
(with router.name=compositeId) on Solr 5.
Quering with "_route_" should not eliminate distributed search overhead?

What is the difference in distribution mechanism between Solr 4 & Solr 5? 
Especially is there any change in filter cache management? 
We see very low hit ratio, with many evictions. Seems that specific shard 
holds entries from other shards (based on _route_ field). Why this can 
happen?

Yes, we have facets, not in all queries. Using regular facets, haven't 
upgraded to Json facets yet.

Appreciate your help,
Esther




From:   Toke Eskildsen <t...@statsbiblioteket.dk>
To:     <solr-user@lucene.apache.org>
Date:   27/08/2015 12:27 PM
Subject:        Re: Solr 5.2.1 versus Solr 4.7.0 performance



On Thu, 2015-08-27 at 11:23 +0300, Esther Goldbraich wrote:
> We are using GC tuning options: Xgcpolicy:gencon , verbose:gc.
> RAM: 64GB
> Solr heap: -Xms512m -Xmx32768m
> Index per server: 500G

Expecting "Your RAM size should be equal to index size"-posts to arrive
in 3, 2, 1...

> Surprisingly, running different setup on same machines, 64 collections / 
1 
> shard per collection gives significantly better results.
> Any ideas? 

Compared to what? If your other scenario is to have 1 collection split
over 64 shards, then the difference boils down to distributed search vs.
single-shard search. There is a non-trivial overhead with doing
distributed search, so if a collection fits well into a single shard
(replicas does not count as distribution here), that is preferable.

Come to think about it, distribution in itself might account for the
difference between 4 & 5 that you are observing. Are you doing faceting
as part of your test?

- Toke Eskildsen, State and University Library, Denmark





Reply via email to