Scrooge Attack: Undervolting ARM Processors for Profit: Practical experience report

Latest ARM processors are approaching the computational power of x86 architectures while consuming much less energy. Consequently, supply follows demand with Amazon EC2, Equinix Metal and Microsoft Azure offering ARM-based instances, while Oracle Cloud Infrastructure is about to add such support. We expect this trend to continue, with an increasing number of cloud providers offering ARM-based cloud instances. ARM processors are more energy-efficient leading to substantial electricity savings for cloud providers. However, a malicious cloud provider could intentionally reduce the CPU voltage to further lower its costs. Running applications malfunction when the undervolting goes below critical thresholds. By avoiding critical voltage regions, a cloud provider can run undervolted instances in a stealthy manner. This practical experience report describes a novel attack scenario: an attack launched by the cloud provider against its users to aggressively reduce the processor voltage for saving energy to the last penny. We call it the Scrooge Attack and show how it could be executed using ARM-based computing instances. We mimic ARM-based cloud instances by deploying our own ARM-based devices using different generations of Raspberry Pi. Using realistic and synthetic workloads, we demonstrate to which degree of aggressiveness the attack is relevant. The attack is unnoticeable by our detection method up to an offset of −50 mV. We show that the attack may even remain completely stealthy for certain workloads. Finally, we propose a set of client-based detection methods that can identify undervolted instances. We support experimental reproducibility and provide instructions to reproduce our results.

I. Introduction

Cloud providers continuosly upgrade their commercial offerings to adapt to market and customer needs. While the vast majority of them offer computing instances based on x86 processors, the availability of ARM-based cloud instances is quickly expanding. ARM processors are increasing their market share of server-grade machines [1]–[8], thanks to additional energy and performance improvements. Before ARM announced its Neoverse [9] microarchitecture, there were no server-grade ARM processors to license. Companies had to customize application-grade ARM processor designs for their server-grade platforms [5], [7]. Recent ARM server-grade processors [1], [3], [7], [8] are based on custom-developed ARMv8 microarchitectures. For example, Amazon [5] deploys ARM-based processors currently shipped in off-the-shelf ARM hardware. Their AWS Graviton processor is essentially a more powerful quad-Raspberry Pi 4B [10]. Scaleway offered instances based on custom-made ARM SoCs with servers smaller than a business card [11]. Table I summarizes a subset of available server-grade ARM processors, supported instruction set architectures (ISA), and providers deploying this hardware. Several generations of ARM processors [1], [2], [5]–[7] are currently available across cloud providers. ARM processors also started reaching into the supercomputing market segment. We expect an increasing availability of ARM Neoverse processors and future server-grade ARM instances to close the performance gap to x86. On the one hand processor manufacturers specify conservative voltage margins due to process variation [12]. On the other hand processors offer different power management mechanisms to adjust frequencies and voltages. While marginal energy savings on a single device appear unimportant, it is of importance at scale, especially since power savings of the cloud infrastructure accumulate for each CPU. The energy footprint of a single execution step (i.e., one single instruction on a processor) is fairly independent of the CPU frequency but dependent on the CPU voltage [13]. Decreasing the CPU voltage below the nominal value to conserve power is called undervolting¹

Notice that Dynamic Voltage and Frequency Scaling (DVFS) differs from undervolting by decreasing frequency as well as voltage.

. Besides energy savings, undervolting directly influences core temperature and can also reduce core aging [14]. Undervolting, however, incurs the risk of introducing soft [15] and hard-errors related to timing violations [16]. These types of errors can be mitigated by carefully analyzing the guardband of processors [17]. In this practical experience report, we consider a scenario where processors supporting a cloud infrastructure are undervolted by an excessively economic and malicious cloud provider (a scrooge §III-A) to profit from additional electricity bill savings, while cloud users (from here on referred to as users) observe similar performance. Unfortunately, undervolting cannot be applied arbitrarily. In fact, it comes at the cost of processor reliability when the supplied voltage is insufficient to drive the processor's frequency. We believe this is a risk that malicious cloud providers are willing to take. For users, undervolting opens up a new attack vector against their cloud applications (see our threat model in §III). The main research questions we address in this work are: RQ1:

What is necessary for a malicious cloud provider in order to pull off a stealthy undervolting strategy?

RQ2:

Does a cloud user have the ability to uncover such an undervolting strategy?

Table I: List of server-grade and mimicking ARM processors with their supported ISA. ‘*’: Used in our evaluation (see §V).

Processor	ISA	Cloud provider
Ampere Altra	ARMv8.2+	Equinix, Oracle
Ampere eMAG 8180	ARMv8	Equinix
AWS Graviton	ARMv8	AWS
AWS Graviton 2	ARMv8.2	AWS
Fujitsu A64FX	ARMv8.2	‐
Huawei Kunpeng 920	ARMv8.2	‐
Marvell ThunderX	ARMv8	Equinix
Marvell ThunderX2	ARMv8.1	Microsoft Azure
NVIDIA Grace	TBA	‐
Broadcom BCM2837(B0)*	ARMv8	‐
Broadcom BCM2711*	ARMv8	‐