Hi,
   Currently, I use 2 servers(2 shards+1 replica) to query 0.46 billion
documents. The average query response time is 1.5s.
   I wish I can simply add a new server into the cluster to improve the
query performance. (still 2 shards + 1 replica)
   But after adding a new server into the cluster, I found the query
performance didn't improve much. The average query response time still is
1.5s.
   Did I miss anything??
   In addition, I'm wondering is there any difference of querying only one
server or query 3 servers roundly?

-Regards-
-Jackie-

* What's your HW setup like, is this data in RAM, on disk, what does
the CPU/disk use etc. look like when you do queries
* What does your data look like, do you have an example
* How does the mapping look like
* How do the queries that you've tried look like
* Have you tried increasing the number of shards, if you have 2 shards
in total with X number of replicas and 3 machines you'll only
distribute your queries to 2 shards, maybe that's your bottleneck, so
increasing replicas/boxes wouldn't help.

Jackie,

It doesn't help to add a new server if no shards/replicas are migrated to
it.  Check manually via the API or look at SPM.
Also, by adding more servers you are adding more CPUs, so resharding may
make sense.

Otis
--
Search Analytics - http://sematext.com/search-analytics/index.html
Scalable Performance Monitoring - http://sematext.com/spm/index.html