Membase Bucket - Performance Concern with multi node setup

I was looking for a distributed key value database as cache for our existing RDBMS project(it lacks performance now..) and found membase as a reliable solution. I started using it membase 1.7. But when I compared performance of membase cache with existing RDBMS database cache by storing and getting large records, iam getting poor performance with membase cache. Testing scenario follows;
I configured servers such a way to form 1 clustor (one master server and one replica server which is created as "join clustor" say for example server1 and server2 - both with membase bucket name "testbucket"). The configuration file is :

When I stored 500 data( say key as "testkey1" ), it always stores in server2(replica server) and when I stored 1000 data ( say key as "testkey2" ), it always stores in server1. I checked it many times by swapping uris ( first server2, then server1) , clearing cache, restarting membase servers. But all are in vein. still 1000+ records always stores in server1 irrespective of configuration settings.

Using existing RDBMS cache:
Step1: Clear all RDBMS cache
Step2: Get(testkey1)- for the first time,as there is no items in RDBMS cache, it will fetch(500 items) from my RDBMS databse.
Step3: Again call Get - as there is item in RDBMS cache, it will fetch from cache. Say time for Getting from cache: 50ms
Step4: Clear all RDBMS cache
Step5: Get(testkey2)- for the first time,as there is no items in RDBMS cache, it will fetch(1000 items) from my RDBMS databse.
Step6: Again call Get - as there is item in RDBMS cache, it will fetch from cache. Say time for Getting from cache: 100ms

Using Membase cache:
Step1: Clear all membase cache
Step2: Get(testkey1)- for the first time,as there is no items in membase cache, it will fetch(500 items) from my RDBMS databse. After that it will store this data to membase cache.
Step3: Again call Get - as there is item in membase cache, it will fetch from cache. Say time for Getting from cache: 100ms
Step4: Clear all membase cache
Step5: Get(testkey2)- for the first time,as there is no items in membase cache, it will fetch(1000 items) from my RDBMS databse. After that it will store this data to membase cache.
Step6: Again call Get - as there is item in membase cache, it will fetch from cache. Say time for Getting from cache: 50ms

In both cases, my perfomance application runs in Server1(master machine)

Although for large records (1000), there is significant performance improvement (over 50%) when using membase cache got, it is noted that for 500, membase cache has poor performance (degradation) compared to RDBMS cache. When I checked the membase console, i found out the reason that testkey2(1000 item key) always stores to server1 and testkey1(500recods) stores in server2(replica) server always. So the reason for much gain of performance may be due to fetching of data from the same machine(server1) that my performance application runs. Because degration happens in fetching data via network ( as my application is in server 1 and 500 data stored in server2). I checked it many times by swapping uris ( first server2, then server1) , clearing cache, restarting membase servers. But all are in vein. still 500+ records always stores in server2 irrespective of configuration settings. So how can I gain or trust high performance in distributed enviornment if I have more number of clustor node and replica nodes? I heard that the internal logic of membase server decides the storing of items in different nodes ( based on which parameters?). Could you please reply?