SSD metrics that matter: Beyond traditional benchmarks

When evaluating SSDs, traditional metrics such as sequential read and write throughput, and random read and write IOPS are often the go-to benchmarks for many customers and system vendors. While these metrics are useful for some comparisons, the criticality of the workloads we run today requires us to look at our storage devices in a different light. In this blog post, we'll explore additional metrics that provide a more comprehensive understanding of SSD performance, especially for specific workloads.
 

IOPS per watt: Efficiency matters

One crucial metric that often gets overlooked is IOPS per watt. This metric measures the number of input/output operations per second (IOPS) that an SSD can perform for each watt of power consumed. In data centers where energy efficiency is paramount, IOPS per watt is a critical factor. High IOPS per watt means that the SSD can deliver high performance while consuming less power, leading to significant cost savings in terms of energy bills and cooling costs. For instance, in graph neural network (GNN) training workloads the Micron 9550 SSD has shown almost double the power efficiency in IOPS/W versus competitors.¹

IOPS per terabyte: Maximizing storage efficiency
 

Another important metric is IOPS per terabyte (IOPS/TB). This metric is particularly relevant when comparing SSDs to nearline hard drives, which offer large capacities but fall short in performance. By assessing IOPS per terabyte, we can determine how efficiently an SSD handles IOPS relative to its storage capacity. This is crucial for customers looking to consolidate multiple hard drives into fewer SSDs, leading to significant savings in space and operational costs. For example, the Micron 6500 ION SSD offers 28% higher capacity than a 24TB HDD² while delivering an astonishing 4,650 times greater random read IOPS per terabyte.

Latency and quality of service (QoS): Consistency and speed
 

Latency and quality of service (QoS) are critical metrics that can significantly impact SSD performance. Latency measures the time it takes for a data request to be processed and returned, while QoS describes the consistency of an SSD's performance. Low latency is essential for applications that require real-time data processing, such as online transaction processing (OLTP) and high-frequency trading. High QoS ensures that the SSD delivers predictable performance, which is crucial for applications like databases and video streaming where performance consistency is vital. The Micron 7500 SSD, for instance, delivers latencies below 1ms for six-nines (99.9999%) latency in mixed, random workloads, ensuring reliable and prompt data delivery. In QoS sensitive applications, like RocksDB, the Micron 7500 is best in class for performance and QoS.³

Terabytes written: Data capacity over time
 

Drive writes per day (DWPD) is a measure of how many times an SSD can be written to its maximum capacity every day for the duration of the drive’s warranty. DWPD is measured assuming fully random workloads, which represent the worst possible condition for a drive due to write amplification (WAF).⁴ DWPD was developed for compute workloads with temporary data that is constantly being updated, meaning the drive is working at the same rate for its lifetime.

However, in storage, the workloads often change over time. Drives fill up quickly early in their lifetime, but access and writes often decline as the drives age . In these cases, DWPD, based on consistent usage, doesn’t match well to an irregular usage pattern. DWPD is not a good metric for high-capacity drives because the particular workload dramatically affects the number of writes the drive can support.

Take, for example, the Micron 6500 ION. At 30.72 TB of capacity and a rating of 0.3 DWPDs, the drive would appear to support a total of 16,819 terabytes written over 5 years. However, because many customers use these drives in sequential fill workloads, the reality is that the drive can support closer to 56,064 terabytes of data written, which is over three times the rated DWPD.

Customers who need high random write drives — such as to house a MySQL TempDB with heavy log writes — sometimes pair their high-capacity solutions with high-endurance solutions such as the Micron XTR. This drive has a smaller capacity by comparison (e.g., 1 or 2TBs of total capacity) but can sustain up to 35 100% 4K random drive writes per day for 5 years.⁵

Furthermore, we are starting to see customers who are looking to consolidate many hard drives with a smaller number of high-capacity SSDs to focus on terabytes written per terabyte per year, which measures the write pressure on the drive while normalizing it for different total capacities.
 

Fill rate: Rebuild times matter

At the beginning of my career, I spent many years as a systems administrator for a large database. The one lesson I learned is that drives fail. The question is not if, but when, so you plan for that. My nightmare scenario wasn’t that a drive failed. It was that multiple drives failed at the same time. Administrators plan for redundancy and failover mechanisms, but every drive failure represents a reduction in reliability. Can we sustain two drive failures? What about three? This is where fill rate comes in. It is a metric that provides a measure to compare rebuild times on SSDs versus hard drives. This is particularly important in scenarios where data recovery and rebuild times are critical, such as in RAID configurations . SSDs generally have faster fill rates compared to hard drives, which means they can rebuild redundancy more quickly in the event of a failure. This reduces downtime and ensures that systems are back up and running swiftly. For example, the 30.72TB Micron 6500 ION SSD, which, thanks to its sequential write speeds of 5GB/s, can fill the drive in a mere 1.7 hours. Compare this to a 28TB HDD⁶ with a sustained transfer rate of 265MB/s which would take 29.4 hours to fill, and you can see that the Micron 6500 ION SSD can fill its larger capacity 17.2 times faster.

Conclusion
 

While traditional metrics like sequential read and write throughput, and random read and write IOPS are important, they don't provide a complete picture of an SSD's performance. By considering additional metrics such as IOPS per watt, IOPS per terabyte, latency and quality of service, terabytes written, and fill rate, we can gain a more comprehensive understanding of an SSD's capabilities. These metrics are crucial for selecting the right SSD for specific workloads, ensuring optimal performance, efficiency and longevity.