Page 34 - EE|Times Europe Magazine - December 2020

34 EE|Times EUROPE

Cloud-Native Processors for a Cloud-Native World

bled over the course of years and multiple
platform generations. Deploying customer
workloads in this heterogeneous infrastruc-
ture can become a bin-packing problem of
nightmarish proportions.
In addition, enterprise-class processors
might have specialized features like Intel’s
AVX-512 that can help improve perfor-
mance on certain applications but do not
have broad applicability. When not used,
they still consume die area and platform
power budget that could have been used for
features with broader applicability. Having
a subset of platforms in a CSP’s data center
that support these specialized features
also means that the CSP must now deal
with a level of heterogeneity while orches-
trating workloads.
Cloud-native processors that are
Figure 3: Normalized plot of performance as a function of the number of containers general-purpose enough to provide leadership
shows how performance falls off for an enterprise-class processor using SMT (blue and performance for the vast majority of cloud
green) versus a cloud-native processor (the Ampere Altra, red). workloads without resorting to costly and
inefficient designs to address specific niche
usage models make life easier for infrastructure teams.
model, which is to maximize revenue by maximizing the number of
users occupying the shared infrastructure while ensuring end-user CONCLUSION
SLAs are not violated. Building out cloud data centers with enterprise-class processors with
Cloud-native processors provide a solution. The higher number features like SMT and Turbo Boost creates a major conflict for CSPs. To
of cores with dedicated execution resources makes it cost-effective maintain profitability, they need to consolidate and maximize resource
for CSPs to rent out each core to a single customer. End users get utilization. Doing so, however, runs the risk that workload contention can
the scalability they want and the predictable performance they expect negatively impact customer SLAs and add unpredictability. The obvious
(Figure 3). The CSPs maximize resource utilization and revenues. solution of making a pair of SMT threads the smallest unit of compute
creates predictable performance but wastes capacity and strands power. In
UNPREDICTABLE PERFORMANCE VARIATIONS addition, Turbo Boost adds performance unpredictability, depending on
Enterprise-class processors with SMT can be extremely effective for the the types of workloads run in multi-tenant environments.
right platform and use case but not for highly scalable and latency- The solution is to select cloud-native processors such as the
sensitive cloud-native workloads. Because SMT threads are not the Ampere Altra. These processors sport an industry-leading core count
same as physical cores, increasing the number of threads, for example, and feature set customized for cloud usage, enabling CSPs to assign
does not provide a corresponding performance increase. A 50% overall resources more effectively and in a predictable manner while improv-
utilization on single-threaded CPUs is not the same as 50% on CPUs ing customer isolation and reducing the attack surface. ■
with SMT.
Turbo Boost adds its own variability. Recall that depending on the Naren Nayak is senior director of application engineering at Ampere
number of cores utilized, the core frequencies are determined by the Computing.
available TDP. Noisy neighbors running
CPU-intensive workloads could consume
higher power, leading to an end-user core
running at different frequencies for a given
workload.
A benchmark run between enterprise-
class processors with SMT and cloud-
native processors demonstrates the effect
(Figure 4). Performance of the enter-
prise-class processor drops because of Turbo
Boost as more cores are used. Once physical
cores are consumed and SMT threads are
used, however, performance drops drastically.
The curve for the cloud-native processor,
which does not support SMT or opportunistic
Turbo Boost, remains linear.
Figure 4: This plot of performance as a function of threads for a 56-core SMT
INFRASTRUCTURE CHALLENGES enterprise-class processor (blue) and 128-core SMT enterprise-class processor (green)
Managing a hyperscale data center is shows performance tapering off as SMT threads are used. The 160-core cloud-native
an enormous job. Facilities can contain processor with single threading (Ampere Altra, red) maintains a linear performance
hundreds of thousands of servers, assem- increase throughout.

DECEMBER 2020 | www.eetimes.eu