On Fri, Apr 1, 2011 at 1:15 PM, Peter Schuller <peter.schul...@infidyne.com> wrote: >> Now, I moved the tokens. I still observe that read latency deteriorated with >> 3 machines vs original one. Replication factor is 1, Cassandra version 0.7.2 >> (didn't have time to upgrade as I need results by this weekend). > > Read *latency* is fully expected to increase if you just add a node. > *Throughput* should increase, unless you have a workload that manages > to be more expensive on RPC than actual reads/writes. > > Latency would only be improved by additional nodes under some significant > load. > > How are you benchmarking? Are you concurrently submitting requests to > all nodes at the same time? Try using stress.py from the Cassandra > tree as a comparison. > > If you're sending one request at a time, there is no expectation at > all of a performance improvement - just a decrease in performance. > > -- > / Peter Schuller >
To be clear on this issue. It does not matter where the tokens start it only matters that they are equally spaced around the token space. So for a 4 node clusters your tokens should either be 1 * ((2^127) / 4) = 42535295865117307932921825928971026432 2 * ((2^127) / 4) = 85070591730234615865843651857942052864 3 * ((2^127) / 4) = 127605887595351923798765477786913079296 4 * ((2^127) / 4) = 170141183460469231731687303715884105728 or 0 * ((2^127) / 4) = 0 1 * ((2^127) / 4) = 42535295865117307932921825928971026432 2 * ((2^127) / 4) = 85070591730234615865843651857942052864 3 * ((2^127) / 4) = 127605887595351923798765477786913079296 If you move one you have to move the rest because the distance between 170141183460469231731687303715884105728 and 0 is 1