Slower Insert Performance in Couchbase Cluster

1
votes

I am new to Couchbase. I was in the process of checking raw performance of INSERTs with various Couchbase setups (single node, 2-node, 5-node etc.) I am using VMs on Windows Azure to do my testing since I do not have enough hardware to test it on my servers. I am using the .NET client library provided by Couchbase to write my C# based test application that does record insertion in a loop. Also, I am not using any bulk insert options (if any exist!)

Here are the results of inserts thus far:

  • Time taken for 10 million records on a hardware server - roughly 10 minutes
  • Time taken for 10 million records on a VM server on Azuere - roughly 21 minutes
  • Time taken for 10 million records on a 2-node VM cluster on Azure 63 minutes

So here are my questions:

  • Is visualization causing the insertion to come down to half of a hardware server?
  • Is it normal for a Couchbase 2 node cluster to take 3 times as much compared to a 1 node Couchbase?

I have followed this article to set up Couchbase cluster in Windows Azure: Couchbase on Azure

1
virtualizationcouchbase
Can you share some code of your test application? What methods did you use for insertion? Where is your client located? May be slowdown is also on client side? - m03geek
Hi, thanks for the response. The same client application is used in both the cases so the client code should not be an issue. If VM1, VM2 are the nodes of my cluster then the client application is run on VM1. - Vinod
I have even "better" distribution in linked question: - 1000% latency for N put requests, where N is max.75% of total requests. - kagali-san

1 Answers

2
votes

Couchbase usally intensivly uses hard disk and cpu for indexing process. So performance degradation in cloud is normal. You can try to use disks with more IOPS or even SSDs to reduce latency.

While your app generates data it also uses cpu, so couchbase performance (on the same VM) can degrade. For more clear test, I think you should run application on separated (dedicated) VM (i.e. VM3). Also performance differs due to latency between VMs. To reduce it, try to locate VMs in one zone, datacenter. I think if you use separate VM for client difference between 1 and 2 node cluster will not be so huge.

In your case performance on 1 node cluster depends mostly on latency between VM and hard disk attached to this VM. In 2 node cluster performance depends mostly on latence between VM1 and VM2 that can be 2-4 times bigger than latency between VM and hard disk. So performance in your cases will be inversely proportional to latences.

performance_1 ~ 1 / HDD_latency_VM1
performance_2 ~ 1 / (HDD_latency_VM1 + Network_latency_2_VM2 + HDD_latency_VM2)