Cloud Storage Throughput Benchmark

Cloud storage throughput benchmarks characterize how fast managed storage services move data. They cover object storage (S3-style buckets), block storage (network-attached volumes), and file storage (managed NFS/SMB). Each has different performance models, and applications stall or overpay when storage is mismatched to the access pattern.

Throughput is rarely a single number. It depends on object or block size, read/write mix, sequential versus random access, concurrency, and whether the service throttles based on provisioned limits or request rate.

What It Measures

Core metrics are aggregate throughput (MB/s or GB/s), IOPS (operations per second), and latency including time-to-first-byte for object reads. Benchmarks report how these scale with parallelism, how they behave at small versus large object/block sizes, and where throttling or burst-credit exhaustion occurs.

Methodology

Load generators such as fio for block and file storage, or parallel object clients for buckets, drive the storage under controlled patterns. Tests sweep block/object size, queue depth or client concurrency, and read/write ratio. For object storage, single-stream throughput is measured first to find per-connection limits, then many parallel streams reveal aggregate scaling. For block volumes, the benchmark distinguishes baseline provisioned performance from burst behavior by sustaining load long enough to exhaust credits. Each configuration runs long enough to reach steady state, and percentile latencies are recorded alongside averages. Caching layers are controlled for by using data sets larger than cache or by issuing cold reads.

How to Interpret Results

Match the result to your workload. Analytics and backups care about large-object sequential throughput and parallelism; databases care about small random-IO IOPS and tail latency. Object storage usually scales throughput by adding parallel connections rather than per-stream speed, so a low single-stream number is normal. For block storage, confirm whether the headline figure is baseline or burst, because burst evaporates under sustained load. Latency percentiles matter more than averages for interactive workloads. Compare against the service's documented provisioned limits to see whether you are throttled or simply under-driving the storage.

Limitations

Results are sensitive to client network bandwidth, region, instance type, and the number of clients, so a benchmark can measure the client rather than the storage. Burst credits, caching, and noisy-neighbor effects make single runs unreliable. Object-storage performance models differ across providers, limiting cross-cloud comparison. Synthetic uniform patterns may not match real hotspots and key distributions, which can trigger per-prefix or per-shard throttling not seen in the benchmark.