For this second report, it's easy: switching from a single query server to a sharded query is going to be slower. Virtual machines add jitter to the performance and response time of the front-end vs the query shards. Distributed search does 2 round-trips for each sharded query. Add these all up and your response time curve flattens out.

Here's how to consider it, using probability arithmetic: suppose the best case is 1 and the worst case is zero, and the mean is .8. If you put two of these measurements in a row, the overall mean becomes 0.8 * 0.8 = 0.64. This is a longer, flatter curve. If a simple search is one round-trip measurement, a distributed search has three measurements in a row. Or, 0.8 cubed = .512. The standard deviation is the flatness of the curve and the fatness of the tail. When you add in the jitter caused by using virtual servers, the standard deviation of the curve increases, making the curve flatter and the long tail fatter. Notice that his best-case query time was faster in 4.0 than with 3.6.1. The core 4.0 data structures are much cleaner and faster. It's the distributed topology that's killing him.

There is no law that says you can't use the indexer/query topology in 4.0. SolrCloud's virtues only kick in after your deployment need several shards.

On 01/17/2013 08:08 AM, Otis Gospodnetic wrote:
Hello,

Here is another one from the other day:
http://search-lucene.com/m/tqmNjXO51B/SolrCloud+Performance+for+High+Query+Volume

Am I the only one seeing people reporting this? :)

Otis
--
Solr & ElasticSearch Support
http://sematext.com/





On Mon, Jan 14, 2013 at 10:55 PM, Otis Gospodnetic <
otis.gospodne...@gmail.com> wrote:

Hi,

I've seen this mentioned on the ML a few times now with the most recent
one being:


http://search-lucene.com/m/mbT4g1fQPr91/?subj=Solr+4+0+upgrade+reduced+performance

Are there any known, good Solr 3.x vs. Solr 4.x benchmarks?

Thanks,
Otis
--
Solr & ElasticSearch Support
http://sematext.com/





Reply via email to