Milvus 2.2 Benchmark Test Report

This report shows the major test results of Milvus 2.2.0. It aims to provide a picture of Milvus 2.2.0 search performance, especially in the capability to scale up and scale out.

We have recently run a benchmark against Milvus 2.2.3 and have the following key findings:

• A 2.5x reduction in search latency
• A 4.5x increase in QPS
• Billion-scale similarity search with little performance degradation
• Linear scalability when using multiple replicas

For details, welcome referring to this whitepaper and related benchmark test code.

Summary

• Comparing with Milvus 2.1, the QPS of Milvus 2.2.0 increases over 48% in cluster mode and over 75% in standalone mode.
• Milvus 2.2.0 has an impressive capability to scale up and scale out:
  • QPS increases linearly when expanding CPU cores from 8 to 32.
  • QPS increases linearly when expanding Querynode replicas from 1 to 8.

Terminology

Test environment

All tests are performed under the following environments.

Hardware environment

Software environment

Deployment scheme

• Milvus instances (standalone or cluster) are deployed via Helm on a Kubernetes cluster based on physical or virtual machines.
• Different tests merely vary in the number of CPU cores, the size of memory, and the number of replicas (worker nodes), which only applies to Milvus clusters.
• Unspecified configurations are identical to default configurations.
• Milvus dependencies (MinIO, Pulsar and Etcd) store data on the local SSD in each node.
• Search requests are sent to the Milvus instances via Milvus GO SDK.

Data sets

The test uses the open-source dataset SIFT (128 dimensions) from ANN-Benchmarks.

Test pipeline

1. Start a Milvus instance by Helm with respective server configurations as listed in each test.
2. Connect to the Milvus instance via Milvus GO SDK and get the corresponding test results.
3. Create a collection.
4. Insert 1 million SIFT vectors. Build an HNSW index and configure the index parameters by setting M to 8 and efConstruction to 200.
5. Load the collection.
6. Search with different concurrent numbers with search parameters nq=1, topk=1, ef=64, the duration of each concurrency is at least 1 hour.

Test results

Milvus 2.2.0 v.s. Milvus 2.1.0

Cluster

queryNode:
  replicas: 1
  resources:
    limits:
      cpu: "12.0"
      memory: 8Gi
    requests:
      cpu: "12.0"
      memory: 8Gi

Search performance

Standalone

standalone:
  replicas: 1
  resources:
    limits:
      cpu: "12.0"
      memory: 16Gi
    requests:
      cpu: "12.0"
      memory: 16Gi

Search performance

Milvus 2.2.0 Scale-up

Expand the CPU cores in one Querynode to check the capability to scale up.

queryNode:
 replicas: 1
 resources:
   limits:
     cpu: "8.0" /"12.0" /"16.0" /"32.0"
     memory: 8Gi
   requests:
     cpu: "8.0" /"12.0" /"16.0" /"32.0"
     memory: 8Gi

Search Performance

Milvus 2.2.0 Scale-out

Expand more replicas with more Querynodes to check the capability to scale out.

queryNode:
  replicas: 1 / 2 / 4 / 8      
  resources:
    limits:
      cpu: "8.0"
      memory: 8Gi
    requests:
      cpu: "8.0"
      memory: 8Gi

What's next

• Try performing Milvus 2.2.0 benchmark tests on your own by referring to this guide, except that you should instead use Milvus 2.2 and Pymilvus 2.2 in this guide.