Methodology
Object storage benchmark methodology
How S3-compatible object storage GET and PUT throughput and small-object latency are measured with MinIO warp from an instance in the same region.
Version 1.0 Updated:
Scope
- S3-compatible object storage reached from a compute instance in the same region
- Large-object GET and PUT throughput (64 MiB objects)
- Small-object GET p99 latency (16 KiB objects)
- S3 API feature checks (multipart upload, presigned URLs, versioning, lifecycle, object lock)
| Tool | Version |
|---|---|
| MinIO warp | Recorded per run |
Setup
A new bucket is created in the region closest to the test instance. The client is the 8 vCPU general-purpose shape from the compute methodology, so that client CPU and network are not the bottleneck. Endpoint URL, region name and TLS settings are recorded.
Workloads
| Workload | warp command |
|---|---|
| Large GET | warp get --obj.size=64MiB --objects=200 --concurrent=32 --duration=5m |
| Large PUT | warp put --obj.size=64MiB --concurrent=32 --duration=5m |
| Small GET | warp get --obj.size=16KiB --objects=10000 --concurrent=16 --duration=5m |
Throughput is reported in MiB/s; latency as p50 and p99 per request.
Feature checks
A script exercises multipart upload, presigned GET and PUT, versioning, lifecycle expiration, bucket policies and object lock. Each feature is recorded as supported, unsupported or behaving differently from Amazon S3, with the error returned.
Repetitions and verification
Each workload runs 3 times on different days. Verification requires a re-run within ±10%.
Limitations
- Throughput is limited by the client instance’s network allowance, which is recorded.
- Results are for one bucket in one region; providers may shard performance per bucket or prefix.