Abstract
In cloud computing, network Denial of Service (DoS) attacks are well studied and defenses have been implemented, but severe DoS attacks on a victim's working memory by a single hostile VM are not well understood. Memory DoS attacks are Denial of Service (or Degradation of Service) attacks caused by contention for hardware memory resources on a cloud server. Despite the strong memory isolation techniques for virtual machines (VMs) enforced by the software virtualization layer in cloud servers, the underlying hardware memory layers are still shared by the VMs and can be exploited by a clever attacker in a hostile VM co-located on the same server as the victim VM, denying the victim the working memory he needs. We first show quantitatively the severity of contention on different memory resources. We then show that a malicious cloud customer can mount low-cost attacks to cause severe performance degradation for a Hadoop distributed application, and 38X delay in response time for an E-commerce website in the Amazon EC2 cloud. Then, we design an effective, new defense against these memory DoS attacks, using a statistical metric to detect their existence and execution throttling to mitigate the attack damage. We achieve this by a novel re-purposing of existing hardware performance counters and duty cycle modulation for security, rather than for improving performance or power consumption. We implement a full prototype on the OpenStack cloud system. Our evaluations show that this defense system can effectively defeat memory DoS attacks with negligible performance overhead.
Introduction: The Gap in Cloud Isolation
In public cloud environments like Amazon EC2 or Microsoft Azure, the multi-tenancy model relies on a hypervisor to ensure that virtual machines (VMs) belonging to different customers remain isolated. While this software layer effectively prevents unauthorized data access, it fails to address the physical reality of shared hardware. While the hypervisor isolates logical memory, it cannot easily isolate the physical hardware paths—such as caches, buses, and memory controllers—used to access that data.
This architectural reality creates a critical vulnerability: a malicious tenant can intentionally create "resource contention." By flooding shared hardware components, an attacker can starve a victim VM of the resources it needs to function. This is not merely a "noisy neighbor" effect; it is a targeted Memory Denial of Service (DoS) attack that can cripple complex distributed applications.
Memory DoS Attack Strategies
The researchers categorize memory DoS attacks into two primary strategies based on the hardware resources they target: storage-based and scheduling-based contention.
1. Storage-based Contention
This strategy targets the Last Level Cache (LLC). An attacker VM performs "LLC cleansing" by filling the shared cache with its own junk data. This forces the victim VM's data out of the cache, requiring the victim to constantly fetch data from the significantly slower main memory (DRAM). This results in a massive performance penalty for the victim's memory-intensive operations.
2. Scheduling-based Contention
This strategy targets the hardware components responsible for moving data, such as the Ring Bus, the Integrated Memory Controller (IMC), and the memory buses. Attackers can use "exotic" atomic operations—like unaligned or uncached memory accesses—to temporarily lock hardware buses. Alternatively, they can flood the IMC with requests, ensuring the victim's legitimate requests are consistently relegated to the back of the scheduling queue.
Mitigation: Repurposing Hardware for Security
To counter these threats, the authors propose a defense system implemented on OpenStack that repurposes existing Intel hardware features for security monitoring and mitigation.
Pseudo-Isolated Reference Sampling
The primary challenge in detection is establishing a performance baseline. To determine if a VM is under attack, the system needs to know how it performs when isolated. Since pausing all other VMs is too costly, the system uses Duty Cycle Modulation—a power-management feature—to briefly throttle all other VMs. This creates a "quasi-isolated" state where the target VM's performance can be sampled.
Detection and Throttling
The system uses Performance Monitoring Units (PMU) to track hardware events like cache misses. It then applies the Kolmogorov-Smirnov (KS) test to compare current performance against the quasi-isolated baseline. If a significant statistical deviation is detected, the system identifies the attacker and applies "electronic handcuffs" by permanently throttling the malicious VM's duty cycle to a fraction of its normal speed (e.g., 1/16th).
Evaluation and Real-World Impact
The research demonstrates that these attacks are not just theoretical. In a real-world test on Amazon EC2, a single attacker VM caused a 3.7x slowdown for a 20-node Hadoop cluster. For a Magento E-commerce site, response latency increased by 38x, while throughput plummeted to just 13% of normal levels. This highlights an "amplification effect" where attacking a single critical node (like a master server) can bottleneck an entire distributed system.
The proposed defense proved highly effective, achieving a 100% true positive detection rate with minimal false positives. Crucially, the persistent monitoring system adds less than 5% performance overhead, making it a practical solution for production cloud environments. Throttling reduced victim performance degradation from 50% down to just 12% during active cache attacks.
Conclusion
Memory DoS attacks represent a distinct and severe threat vector that bypasses traditional software-based isolation. By repurposing hardware features intended for power management and performance debugging, cloud providers can implement robust, low-overhead defenses. Future work will look toward adapting these techniques for non-Intel architectures and developing strategies to counter attackers who attempt to "pulse" their contention to evade statistical detection.
Summary updated