The Art of CPU-Pinning: Evaluating and Improving the Performance of Virtualization and Containerization Platforms

Abstract

Cloud providers offer a variety of execution platforms in form of bare-metal, VM, and containers. However, due to the pros and cons of each execution platform, choosing the appropriate platform for a specific cloud-based application has become a challenge for solution architects. The possibility to combine these platforms (e.g., deploying containers within VMs) offers new capacities that makes the challenge even further complicated. However, there is a little study in the literature on the pros and cons of deploying different application types on various execution platforms. In particular, evaluation of diverse hardware configurations and different CPU provisioning methods, such as CPU pinning, have not been sufficiently studied in the literature. In this work, the performance overhead of container, VM, and bare-metal execution platforms are measured and analyzed for four categories of real-world applications, namely video processing, parallel processing (MPI), web processing, and No-SQL, respectively representing CPU intensive, parallel processing, and two IO intensive processes. Our analyses reveal a set of interesting and sometimes counterintuitive findings that can be used as best practices by the solution architects to efficiently deploy cloud-based applications. Here are some notable mentions: (A) Under specific circumstances, containers can impose a higher overhead than VMs; (B) Containers on top of VMs can mitigate the overhead of VMs for certain applications; (C) Containers with a large number of cores impose a lower overhead than those with a few cores.

Characterizing network performance of single-node large-scale container deployments

“…This performance trend is due to two reasons: first, many containers share a core with a high ratio, which naturally causes them to get a smaller slice of the CPU. Second, the number of interrupts and context switches on the CPU is high under the high ratio that incurs a high management overhead [2], [20]. Fig.…”

“…For example, pinning communicating containers to different cores has increased OVS throughput around 21.2% compared to the scenario without pinning. We suspect that such improvement is due to the reduced overhead originating from cache misses that follows the occurrence of context switches when a container resumes on a different core [20]. Pinning communicating containers to the same core, on other hand, can drop the total throughput depending on the driver (e.g., up to 13.2% for a bridge interface).…”

To Pin or Not to Pin: Asserting the Scalability of QEMU Parallel Implementation

“…In the end, the execution times of our program subset without and with pinning remain similar and for some programs it even appears to be a little slower when pinning vCPUs. According to [25], CPU-bound applications, such as the PARSEC ones, do not improve the performance when pinning virtual machines. The current Linux scheduler attaches a thread to a CPU as long as it does not degrade performance as estimated at run time.…”

Efficiency in the Serverless Cloud Computing Paradigm: A Survey Study

“…Due to including the whole operating system and application stack in the VM image, in general, VMs suffer from a high memory and storage footprints. In addition, VMs introduce a high startup delay [32] to boot up, hence, they are not a perfect fit for frequent starts and terminations inherent to the serverless functions [11]. As such, VMs are usually employed as the underlying platform of the serverless frameworks, rather than an isolator of each individual function.…”