Page 32 - EE|Times Europe Magazine - December 2020

32 EE|Times EUROPE

MARKET & TECHNOLOGY TRENDS
Cloud-Native Processors for a Cloud-Native

World

By Naren Nayak

he cloud computing market is skyrocketing, with some analysts Enterprise-class processors
expecting it to reach nearly US$761 billion by 2027, an annual Until relatively recently, applications were primarily written as
growth rate of 18.6%. This surge is driven by a veritable laun- monolithic blocks of code meant to be run in dedicated local IT envi-
T dry list of potential benefits. The cloud can enhance utilization ronments. Running them required increasingly powerful CPUs with
and reduce costs through shared infrastructure for cloud service vast amounts of memory. Such applications relied on micro-
providers (CSPs). It can offer a pay-as-you-go model with elasticity that architectural innovations such as higher CPU frequencies, larger
lets end users spin up and tear down compute resources as needed. The caches, and new instructions for specific usages.
decentralized infrastructure can enhance resiliency and ensure avail- The slowing down of Dennard scaling in the 2000s effectively put an
ability for enterprises. Plus, the massive consolidation and economies end to frequency scaling as the primary mechanism for higher perfor-
of scale can provide great savings for CSPs. mance, and multicore CPUs
That is the theory, anyway. Reality can be very different. By making an informed became commonplace. The
Many of the benefits of the cloud come at the cost of running emphasis for enterprise-
workloads in a shared or multi-tenant environment. Depending on the choice of platform class processors, however, has
exact infrastructure, this environment can suffer from several issues, components, it is still been on high-capacity
including: cores rather than scaling
• Unpredictable performance variations (the “noisy neighbor” possible to both support out to larger core counts.
problem) Enterprise-class cores are
• Nonlinear system scalability a profitable cloud designed to run computation-
• Complex, costly, and/or time-consuming infrastructure infrastructure and ally intensive applications
management but don’t necessarily lend
These issues highlight the central conflict for CSPs: Depending deliver the performance themselves well to shared
on the infrastructure, the only way to meet service-level agreements cloud infrastructure.
(SLAs) may be to assign dedicated hardware to each customer, but that customers expect. Catering to the types of
doing so can undercut the economies of the cloud business model. By workloads that run on them,
making an informed choice of platform components such as CPUs (the many enterprise-class processors sport features like simultaneous multi-
most expensive component of the platform bill of materials), it is pos- threading (SMT) and Turbo Boost. SMT enables each physical core to be
sible to both support a profitable cloud infrastructure and deliver the split into two threads, represented as “logical threads” that are capable
performance that customers expect. Let’s take a closer look. of simultaneously executing separate instruction sequences. It is a
technique that can more effectively increase utilization of the execution
DATA CENTER PROCESSOR EVOLUTION resources of an inefficient CPU, but its value is decreased as the system is IMAGE: SHUTTERSTOCK
Before we can discuss the issues, we need a quick review of data center optimized and as generated code better utilizes resources. SMT can also
processor architectures. Data center processors can be roughly divided create more problems than it solves, especially in the cloud.
into two types: enterprise-class processors and cloud-native processors. Turbo Boost is a technology that automatically runs cores faster than

DECEMBER 2020 | www.eetimes.eu