Details
-
Bug
-
Resolution: Fixed
-
Critical
-
7.6.4
-
Operating System : Microsoft Windows Server 2022
Couchbase Enterprise Edition build 7.6.2-3720
-
Untriaged
-
Windows 64-bit
-
-
0
-
Yes
Description
Steps taken for this issue
- Created a 2 node cluster - 1 kv, 1 fts node
- Created a magma bucket and loaded 10M docs onto it (4096 dims)
- Created a vector index - definition attached
The above set of steps was carried out on 16 vCPU machines and 8 vCPU machines in windows. For comparison, the same was also executed on Linux systems.
For 16 vCPU systems - c6a.4xlarge EC2 instances were used
For 8 vCPU systems - t3.2xlarge EC2 instances were used.
The same random data was used in all the clusters. It was loaded onto one of the clusters and was replicated onto all the others using XDCR.
Observed the indexing stats at this point of time
Listing down the observations
- In windows systems the sysproc CPU utilization is around 200% regardless of the number of vCPUs.
- In linux machines, the sysproc CPU utilization goes to almost 1600% in a 16 vCPU node cluster and 800% in a 8 vCPU node cluster
- In Linux clusters, overall CPU utilization always oscillates around 80-100% during indexing indicating good resource utilization
- In Windows clusters, the resource utilization is very low.
Windows - 16 vCPU machines
Windows - 8vCPU machines
Linux - 16 vCPU machines
Linux - 8 vCPU machines