TEST SETUP AND METHODOLOGY
In testing all enterprise drives we focus on long-term stability. In doing so, we stress products not only to their maximum rates, but also with workloads suited to enterprise environments. We use many off-the-shelf tests to determine performance, but we also have specialized tests to explore specific behaviors we encounter. With enterprise drives, you will see that we do not focus on many consumer-level use-cases.
When testing SSDs, the drive is purged and then preconditioned into a steady state before capturing its performance results. We also disable all write caching on the DUT when possible, this ensures consistent results that are compliant with SNIA standards. Our hope is that we present tangible results that provide relevant information to the buying public.
|PC CHASSIS:||Thermaltake Urban T81|
|MOTHERBOARD:||ASRock X99 WS-E|
|CPU:||Intel Xeon E5-2690 v3|
|CPU COOLER:||Thermaltake Water 3.0 Ultimate|
|POWER SUPPLY:||Thermaltake Toughpower 1500W Gold|
|GRAPHICS:||MSI GT 720 |
|SYSTEM COOLING:||be quiet! Silent Wings 2|
|MEMORY:||Kingston ValueRAM DDR4 2400MHz ECC 64GB|
|STORAGE:||Crucial MX200 500GB|
|OS:||Windows Server 2012 R2|
This Test Bench build was the result of some great relationships and purchase; our appreciation goes to those who jumped in specifically to help the cause. Key contributors to this build are our friends at ASRock for the motherboard and CPU, be quiet! for the cooling fans, and Thermaltake for the case. We have detailed all components in the table below and they are all linked should you wish to make a duplicate of our system as so many seem to do, or check out the price of any single component. As always, we appreciate your support in any purchase through our links!
The Storage Networking Industry Association has an entire industry accepted performance test specification for solid state storage devices. Some of the tests are complicated to perform, but they allow us to look at some important performance metrics in a standard, objective way.
SNIA’s Performance Test Specification (PTS) includes IOPS testing, but it is much more comprehensive than just running 4KB writes with Iometer. SNIA testing is more like a marathon than a sprint. In total, there are 25 rounds of tests, each lasting 56 minutes. Each round consists of 8 different block sizes (512 bytes through 1MB) and 7 different access patterns (100% reads to 100% writes). After 25 rounds are finished (just a bit longer than 23 hours), we record the average performance of 4 rounds after we enter steady state.
- Purge: Secure Erase, Format Unit, or vender specific
- Preconditioning: 2x capacity fill with 128K sequential writes
- Each round is composed of .5K, 4K, 8K, 16K, 32K, 64K, 128K, and 1MB accesses
- Each access size is run at 100%, 95%, 65%, 50%, 35%, 5%, and 0% Read/Write Mixes, each for one minute.
- The test is composed of 25 rounds (one round takes 56 minutes, 25 rounds = 1,400 minutes)
Unlike our other performance tests, the SNIA tests only last for a relatively short period of time each (1 minute), but they cover many more access patterns and transfer sizes. All tests were done at a QD of 32. Here the DC400 delivered somewhat underperforming results. It maxed at 78K IOPS at 100% 4K read and almost 6K IOPS at 100% 4K write for both over provisioned and non-over provisioned testing. In the 50% mix we see that over provisioning does indeed help, but not by that much.
To specifically measure latency, we use a series of 512b, 4K, and 8K measurements. At each block size, latency is measured for 100% read, 65% read/35% write, and 100% write/0% read mixes.
As far as average latency results go, the DC400 has average results for 100% reads, but lag slightly behind the competition in both mixed and 100% write workloads. The 960GB model averages about 0.4ms latency during 4K reads, nearly 3ms latency during a 65/35% read/write mix and 5.5ms latency during 100% 4K writes. 512b results nearly mirror 4K and at 8K the latency results are a bit more than double. When provisioned to 800GB, the DC400’s read and write latencies are nearly the same, however, during mixed workloads, the extra over-provisioning helps to cut average latency down to a fraction! At 4K it averages 0.8ms during the 65/35% read/write mix and 8K is simply double that.
Moving onto maximum latency, we can see that the 960GB model spiked to about 8.5-9.5ms during 100% reads while the 800GB model ranged from 6.9-7.4ms. During the mixed workload, the 960GB model peaked at 22-26ms and the 800GB model ranged from 14-18ms. Finally, during 100% writes the 960GB model spiked at 21-32ms while the 800GB model spiked to 15-25ms. All results are in line with what we would expect to see from a read focused product so far.