Follow Datanami:
December 2, 2021

AWS Adds a Little More Nitro to Its SSDs

Looking to serve data to your cloud applications and databases a little faster? Then you might be interested in the new Nitro SSDs that AWS announced this week at re:Invent. Sure, they sound fast, but the Nitro SSDs are no ordinary flash drives.

One of the secret recipes that Amazon Web Services has whipped up that enables it to scale its customers’ workloads to such dizzying heights is its custom-made Nitro chips. These Nitro chips represent AWS’s first foray into the world of custom silicon, and they form the basis for the Nitro controllers that are the backbone of EC2 today.

“Nitro is the reason that AWS got started building its own chips,” Peter DeSantis, the general manager for EC2, said during a keynote address at the re:Invent conference Wednesday afternoon. “And it remains one of the most important reasons why EC2 provides the best performance and security in the cloud.”

According to DeSantis, every EC2 server has a Nitro controller, which embeds the firmware that AWS uses to connect any type of processor (X86, ARM, AMD, Mac, etc.) into its global network. The Nitro controller provides the consistency that customers’ workloads demand, he said.

“The Nitro controller runs all the AWS code that turns that server into an EC2 instance, and there are a number of benefits to this approach,” he said. “By doing all the network and storage virtualization in the Nitro controller, we also reduce variability and avoid interfering with customers workloads, which is first and foremost. And these last two benefits—supporting any type of hardware and improving performance—sounds an awful lot like our problems with SSDs. So it probably won’t surprise you to hear that we build a Nitro SSD.”

The problem with SSDs is that every SSD manufacturer uses its own custom Flash Translation Layer, or FTL. In fact, some SSD manufacturers use multiple FTLs within their SSD lines. FTLs are complicated bits of software that essentially function as mini-databases to ensure the NAND cells look just like random I/O media to the application.

“They all provide generally the same API and they all do a good job with an average case, but our experience over the years is each one has unpredictable and idiosyncratic behaviors,” DeSantis said. “For example, garbage collection can kick in at an unexpected stime and cause I/O requests to stall. And these unexpected behaviors make it really difficult when you’re trying to provide consistent performance. And they make it really hard to run certain workloads, like databases.”

By developing its own Nitro-based FTLs, AWS thinks it has solved the performance issues that have prevented some of the most demanding data workloads from reaping the benefits of SSDs.

“We launched the Nitro SSDs, the NVMe drives ,where we essentially rewrote the whole FTL,” Raj Pai, vice president of EC2 product management told Datanami in a briefing at re:Invent. “We own the firmware. We own the logical mappings of the storage on those devices. By building the whole stack and tightly integration that in our whole system, we’re able to achieve 60% reduction in latency from what you can get with OEM drives, and 75% less jitter [or latency variability] in terms of performance, which allows our customer to reduce the number of nodes they have running their relational databases or NoSQL or analytics.”

Earlier this year, AWS launched the Nitro SSDs as the underlying storage in the io2 Block Express EBS volume. This EC2 instance is designed for block storage workloads that demand lots of IOPS, such as relational databases. It also offered Nitro SSDs in the I4i instance, which is powered by Intel processors.

The announcement at re:Invent yesterday represents the second generation of Nitro SSDs. AWS unveiled Nitro SSDs in two new EC2 instance types this week, including the Im4gn and Is4gen instances, both which support up to 30TB of NVMe storage and feature the Graviton2 processor.

Im4gn instances are good for applications that require large amounts of dense SSD storage and high compute performance, but are not especially memory intensive, such as social games, session storage, chatbots, and search engines AWS says.

“Is4gen instances are a great fit for applications that do large amounts of random I/O to large amounts of SSD storage,” AWS Chief Evangelist Jeff Barr wrote in an AWS blog on Tuesday. “This includes shared file systems, stream processing, social media monitoring, and streaming platforms, all of which can use the increased storage density to retain more data locally.”

“We’re really excited about performance benefits of Nitro SSD,” DeSantis said. “But if we’re going to talk about the big investment we’re making in workloads like databases, we have to talk about Graviton.”

Hyperscalers like AWS traditionally have been known as large buyers of commodity gear. But as we’ve seen with the Nitro chip and its follow-ons, the Graviton, Inferentia, and Trainium chips, it’s clear that AWS has discovered that the key to delivering superior performance for customers is by developing their own hardware.

Related Items:

AWS Rocks with New Analytics, AI Services at re:Invent

AWS Leans on Custom Silicon for Processing Advantage