Slower benchmark with Sidekick

[choyuansu]

I'm using minio/warp to benchmark my MinIO deployment. I tried connecting to the storage nodes directly vs. using Sidekick, and the results from using Sidekick is much slower compared to connecting directly (1200 MiB/s vs 500 MiB/s).

Expected Behavior

I expect the benchmarks to have similar speed whether I using Sidekick or not.

Current Behavior

Using Sidekick results in slower speed.

Possible Solution

No idea.

Steps to Reproduce (for bugs)

1. Assume 4 storage nodes minio{1...4}.example.com:9000.
2. Benchmark w/o using Sidekick:

   warp mixed --host=minio{1...4}.example.com:9000 --access-key=ACCESS_KEY --secret-key=SECRET_KEY --autoterm

3. Benchmark using Sidekick:

   warp mixed --host=localhost:8000 --access-key=ACCESS_KEY --secret-key=SECRET_KEY --autoterm

Context

I'm trying to measure the throughput for my MinIO deployment.

Your Environment

• Version used (sidekick version): 2.0.0
• Operating System and version (uname -a): Ubuntu 20.04

• Sidekick is run using a container

  # docker-compose.yaml
  services:
  sidekick:
    image: "minio/sidekick:v2.0.0"
    ports:
      - 8000:8000
    command: --health-path /minio/health/ready --address :8000 http://minio{1...4}.example.com:9000
  
    extra_hosts:
      - minio1.example.com:x.x.x.x
      - minio2.example.com:x.x.x.x
      - minio3.example.com:x.x.x.x
      - minio4.example.com:x.x.x.x

[klauspost]

(followup on slack)

After inspecting the code, seems like sidekick is using round-robin load balancing by default, and warp balances based on the number of requests for each node. After setting warp to use round-robin, it produced similar results.

Would be fixed by #16