In my previous blog about the Micron 7450 NVMe SSD, I noted that Micron has close relationships throughout the industry that enable us to understand data center workload challenges and requirements, such as latency. In this blog, I take a deeper look at how the Micron 7450 NVMe SSD delivers low and consistent latency benefits. Let’s start with what’s behind recent data center innovation and how Micron is helping.
Four Drivers for Data Center Storage Innovation
There are four main areas where Micron is driving storage innovation:
- Simplifying the Transition to PCIe and NVMe: Most of the customers I talk with have already deployed PCIe/NVMe in some way. NVMe SSDs span boot, main data storage and acceleration with different form factors (M.2 for boot or main data storage, E1.S for main data storage and U.3 for deployed systems) and endurance ratings (read-centric and mixed-use endurance).1
- Storage Density: When we increase storage density (by increasing the SSD capacity on small form factors, like 7.68TB on an E1.S SSD), we yield better data center space utilization and improved efficiency.2
- Security: As data volume and its diversity grow, the industry has seen an increase in the number and sophistication of attacks. Hardware encryption on SSDs has gained more adoption as administrators look to strengthen their security posture, which often includes using Micron security innovations.3
- Low and Consistent Latency: From the largest hyperscalers to emerging data center operators, hosting multiple workloads on a single system (virtualization or containerization)4 is standard practice. These workloads can generate millions of transactions every second with many sharing the same physical storage. Micron designed the 7450 NVMe SSD to meet this demand. The Micron 7450 is an SSD that manages these millions of transactions and does it while delivering low and consistent latency – often called great quality of service (QoS).5
We should remember that many data center applications are real time – they use, process, and respond to data that must be delivered quickly and consistently. Data delivery delays or interruptions can be detrimental to many of these workloads.
One example of a latency-sensitive workload is live streaming. When streaming a live event, the best user experience relies on the streaming service providing low and consistent stream latency. Failure to do so may adversely affect the stream’s quality -- the stream may lag, show buffering delay or lose frames. Consider the example of live sports. Low latency streaming makes the streamed event more lifelike – almost like being there.6 Live streaming isn’t the only latency-sensitive data center workload. Some workloads are more “behind the scenes.” They also rely on fast, consistent data delivery.
Real-time analytics is the practice of extracting insights from data to make better and timely decisions, and data delays can limit its effectiveness. When data is collected from multiple sources, analytics may have to wait for the slowest source – whether that is a single server or a rack of them. When storage responds quickly and consistently, wait times are minimized and insights are delivered faster. Modern data centers have also migrated from single workloads running on bare-metal servers to containerization and virtualization deployments where single servers host multiple applications and workloads, sharing the same physical server resources.
These cloud and virtualized (multitenant) workloads also rely on low, consistent latency. They are highly parallel, multitenant workloads that share the same underlying physical resources (storage, CPU, memory, and networking). When storage responds quickly and consistently, CPU resources can wait less, making it possible to host more workloads or support more virtual machines while providing a more predictable pattern of performance. While storage throughput clearly matters, maintaining high throughput with low and consistent access is critical for data centers to meet their performance goals and service level agreements.
The Fundamentals and Impact of Latency
Quality of Service (QoS) is a metric that describes application latency consistency in an SSD. There are three essential components to measuring QoS: threshold, percentage (of IOs), and IO type:
- A threshold: The response time needed for each access (latency, often measured in milliseconds).
- A percentage: The percentage of all the accesses that are under the threshold. This is often expressed in terms of a percentage of all IOs, like 99.9999% and is often referred to by the
number of nines in this number – 99.99% is “four-nines” and 99.9999% is “six-nines.”
- An IO type: The type of data IO being measured. An IO could be a read, a write, or some combination of the two.
Great QoS Is Essential to Data Center Workloads
Data center and cloud workloads have an insatiable hunger for fast, consistent read performance. For example, on Black Friday 2021, 88 million buyers spent $8.9 billion US (total) in online purchases, with Amazon accounting for 17.7% of those7. These online customers expect their transactions to be completed quickly and consistently (great QoS) so they can move on with what’s next on their “to do” list.
The Micron 7450 SSD addresses QoS needs with industry-leading, 99.9999% mixed-workload read latencies under 2 milliseconds (ms) while still delivering hundreds of thousands of IOPS.8 In more read-centric workloads, it can deliver up to 1 million IOPS (complete performance information is available in the Micron 7450 SSD Product Brief). As we’ll see momentarily, this lower latency enables rack-scale applications to deliver more work than they could with many other drives available today.
The Micron 7450 NVMe SSD Delivers Great QoS
Satisfying these application demands isn’t easy. There are real challenges to achieving low and consistent workload latency in complex, scalable environments. Meeting this challenge was a big focus for us in the development of the Micron 7450 NVMe SSD.
Sub-2ms 99.9999% SSD Read Latency in Mixed, Random Workloads
Let’s see what great QoS looks like. Figure 1 below shows a mixed workload (a 4KB transfer size, 100% random placement with 90% read and 10% write) run on an individual Micron 7450 SSD at six-nines (99.9999%) read latency. Figure 2 shows similar results with the write portion increased to 30%. Figures 1 and 2 show read latency in mixed workloads because write latency can be affected by write concatenation, bifurcation, and caching at the operating system, file system, or application layer (as noted in section 4 of this SNIA presentation).
This is pretty amazing. Unlike many previous SSDs, the Micron 7450 SSD delivers read latencies at or below 2ms — and remains there — for six-nines latency in mixed, random workloads and common queue depths (QDs).9 Based on Micron lab testing, this low, consistent latency can improve performance in databases like Microsoft SQL Server, Oracle, MySQL, RocksDB (a good representation of cloud workloads), Cassandra, and Aerospike, among others.
What Can Happen With Great QoS?
When a data center SSD shows great QoS at a low threshold, the result is clear – more read requests will be below the threshold.
To illustrate the application-level benefits of great QoS, we compared the Micron 7450 NVMe SSD to another mainstream NVMe SSD. We used RocksDB, a performance-focused key-value store (commonly used in latency-sensitive, user-facing applications like storing viewing history and spam detection, among others). When we set a common application delivery threshold, we found that the Micron 7450 SSD reaches 95% higher performance.
When we look at great QoS at rack-scale, its benefits are much clearer than one might assume when looking at QoS data for a single SSD. But how much better? Is there any real difference between, say, 99.9999% (six-nines) and 99% (two-nines) when we look at a rack of servers? There is, and we can calculate the impact.
First, we’ll look at the number of read IOPS possible with a rack full of E1.S Micron 7450 NVMe SSDs. Then, we’ll look at the number of those IOPS that are outside of a 2ms threshold for a different number of nines.
We’ll install 32 SSDs in a 1U server10 and load 38 of these servers in a rack.11 Our E1.S Micron 7450 NVMe SSDs can generate 1 million 4K random read IOPS. When we install 32 of these SSDs per server and load 38 servers per rack, the hypothetical maximum number IOPS would be approximately 32,000,000 per server (or about 1,200,000,000 per rack -- 1 million IOPS per SSD x 32 SSDs per server x 38 servers per rack).
If we start the rack-scale discussion at 99% (two-nines) 2ms QoS read latency, our rack of servers will show about twelve million reads that are outside of our 2ms threshold (1% of the reads will be outside of 2ms). But at six-nines (99.9999%) we are much lower with an approximated average near 1,200 reads outside of our 2ms threshold.
|Read IOPS per SSD||SSD's per Server||Servers per Rack||2ms QoS % Value||Reads outside of 2ms (approx.)|
|1 Million||32||38|| 99%
The rack-scale picture really is different. For our 2ms threshold QoS – 12 million reads outside of our threshold (two-nines), or just twelve hundred (six-nines)? This is a theoretical example, so actual workload-level differences will vary. But low and consistent latency benefits most data center workloads and for those that are particularly latency-sensitive, it is essential.
The Micron 7450 SSD Meets These Needs Head-On, Delivering Low and Consistent Latency
Micron built the 7450 SSD with the most advanced, 176-layer NAND, the world’s most advanced, mass production NAND in the world today, and combined it with our advanced controller and firmware to deliver incredible results.12
But the biggest advantage is our ability to integrate these innovative technologies quickly into our SSDs, which enables more users to have faster access to these innovations.
How to learn more
The Micron 7450 SSD performance is designed to deliver impressive QoS for various data center workloads and shows substantial benefit to complex workloads.
Learn more at the 7450 page on Micron.com and contact your sales representative to get them in your lab – and then into production!
1. The Micron 7450 SSD is available in M.2, E1.S, and U.3 form factors in read-mostly and mixed-use designs. This simplifies system design by enabling a single SSD architecture to satisfy boot, main data storage, and acceleration (caching).
2. Increased SSD capacity means fewer are needed to store the same amount of data. Fewer SSDs of the same form factor need fewer servers to house them and less space to mount those servers. For additional information on the E1.S form factor, see https://www.snia.org/forums/cmsi/knowledge/formfactors. Note: For all capacity statements, formatted capacity will be less.
3. No hardware, software or system can provide absolute security under all conditions. Micron assumes no liability for lost, stolen or corrupted data arising from the use of any Micron products, including those products that incorporate any of the mentioned security features.
4. See https://www.gartner.com/en/information-technology/glossary/virtualization for additional details
5. See https://www.snia.org/educational-library/storage-quality-service-enterprise-workloads-2014 for additional background on applying qualify of service to storage.
6. Additional information is available here: https://www.dacast.com/blog/best-low-latency-video-streaming-solution/#:~:text=Low%20latency%20streaming%20is%20especially%20important%20for%20certain,thing%20to%20attending%20your%20event%20in%20real%20life
8. 2ms latency is a common data center workload latency requirement.
Microsoft notes that “…workloads that need a fast response from the storage layer (1-2 milliseconds in average) should use Business Critical tier…” for Azure SQL Database and Azure SQL Managed Instances: https://docs.microsoft.com/en-us/azure/azure-sql/database/service-tier-business-critical?view=azuresql
IBM also notes a 2ms heartbeat requirement in their high availability requirements: https://www.ibm.com/docs/en/qsip/7.4?topic=planning-link-bandwidth-latency
9. Up to queue depth = 32 for 4KB, 100% random, 70% read workload, based on workload data collected by Micron Engineering
10. Example server: https://www.supermicro.com/en/products/system/1U/1029/SSG-1029P-NES32R.cfm
11. Common server racks hold up to 42 1U servers: https://www.tripplite.com/42u-smartrack-standard-depth-server-rack-enclosure-cabinet-doors-side-panels~SR42UB with additional rack space accommodating switches and other devices.
12. Based on similar use SSDs with NVMe available on the open market as of the date of this document’s publication. The Micron 7450 SSD offers a broader range of form factors combined with industry-leading Micron 176-layer NAND.