A question for result benchmark

I have a result benchmark

When I increase data node and replicas, term latency increase and throught put decrease. Do my config have a problem?
This is my config es master so same es data

cluster.name: cluster-es
node.name: node-xxx
network.host: ["xxx.90.33.22"]
http.port: 9200
discovery.zen.ping.unicast.hosts: ["xxx.90.33.18"]
discovery.zen.minimum_master_nodes: 2
thread_pool.search.queue_size: 1000000
thread_pool.bulk.queue_size: 1000000
indices.breaker.fielddata.limit: 75%

Hi @phucvinh52,

I don't know anything about your benchmark setup (e.g. how did you invoke Rally? On which hardware did you run Rally? How was Rally connected to your cluster?) but there are a lot of throughput numbers missing in the first case (3 data nodes, no replicas). This is an indication for error responses (unfortunately, we don't show the error rate in the summary report yet which would make it very obvious) .

  • Did you check the log files of Elasticsearch and Rally for any errors?
  • Did you specify all your data nodes in the --target-hosts parameter?

Also: These huge queue sizes do more harm than good. You should reduce them significantly (just use the default values). If your queues are constantly full this means that you are throwing more load at the system than it can handle and you should rather increase your capacity (e.g. more nodes, more RAM, more CPU - depends on what you're bottlenecked) than increasing the queue size as the latter just hides the actual problems.

Daniel

1 Like

This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.