Best Practices Cont...
Why You Need a Reliable Storage Benchmark
What's this
2. Confirm enterprise-class workloads match your applications.
Benchmark tests that address streaming reads and writes as well as random read/write IOPS are fairly common, and this is certainly a good starting point. But also consider factors such as query and update operations. These types of workloads are common in Online Transaction Processing (OLTP), email, and database applications. It’s critical to understand your target applications and the ways in which they interact with the storage subsystem. If the benchmarks being used don’t reflect such usage, then the usefulness of those tests for your needs is questionable.
3. Confirm that the test evaluates and reports on responsiveness.
There is more to storage performance than IOPS. Because storage sub-systems are likely to juggle multiple types of operations concurrently, IT decision makers wanting the whole story should also examine average response times as I/O operations and workloads change. This will help indicate sustained consistency of performance. Ideally, the trend line that emerges during benchmarking should look fairly flat. When spikes appear, either up or down, as workloads change, it’s obvious that the storage solution is performing inconsistently.

4. Know the test duration.
Some storage benchmarks only run for a minute or two. As mentioned earlier, this makes for a glaring disconnect between test scenarios and the real world. Business storage subsystems often run around the clock. Even a one-hour test, while providing a general idea of performance, can’t offer a genuine proof of long-term consistency. Four-hour tests are good; eight-hour tests are even better.
5. Ensure that the performance test is repeatable, reproducible, and uses consistent test metrics.
This might seem to go without saying, but many test labs may be rather informally run, lacking consistency in factors such as BIOS configuration, test parameters, and environmental conditions. Everything should be documented, from the driver version for each applicable component to pertinent temperature readings from drives. As with any well-designed science experiment, only this level of detail can help ensure that test results may be independently repeated and verified with a sufficient level of accuracy.
6. Make it public.
The auditor should have nothing to hide and no reason to hide it. Similarly, vendors should be willing to make their results public so that buyers can trust the merchandise inside and out. If you request benchmark results but find either party unwilling to supply them, then either there’s a dangerous dearth of information or someone is hiding something.