TR-2019-12
CLOUD RESOURCE MANAGEMENT FOR BURSTY, REAL-TIME WORKLOADS
Hai Nguyen. 15 July, 2019.
Communicated by Andrew Chien.
Abstract
Today’s cloud resource offerings provide no guarantees for resource allocation, so bursty, real-time applications must reserve, and pay for resources they do not use - to achieve real- time guarantees. We propose Real-time Serverless, a new type of cloud resource with a new service-level objective – guaranteed allocation rate for transient cloud function invocations. This guarantee enables timely resource allocation, enabling applications to achieve real-time performance efficiently with good resource utilization and cost. With a simple burst model, we study real-time serverless analytically, exploring its effect on application quality, guaran- tees, and cost. Next, we use simulation to explore a statistical variety of bursts and higher loads (multi-application), to study the benefits of Real-time Serverless for applications.In both analytic and simulation studies, the availability of real-time serverless service enables bursty, real-time applications to achieve guaranteed high quality at reasonable cost. Specifically, for a desired application quality, the required allocation rate can be determined. Real-time serverless also enables applications to match resources to demands and thereby minimize the cost. For example, for duty factors from 0.025 to 0.25, the value per unit resource of Real-time Serverless instances to a bursty application can be 4x higher than the traditional. Further results show the robustness of real-time serverless with bursty, real-time applications over a wide range of properties. Multiple applications can share a real-time serverless pool efficiently, supporting duty factor increases of 25x with only a 1.6x increase in allocation rate (provider resource cost), revealing high resource management flexibility and economic potential to cloud providers.
We present a case study of a video-based traffic monitoring application. Despite more complex burst statistics, real-time serverless delivers major benefits for application quality and cost. For this application, real-time serverless instances are worth nearly 16x per unit resource, when compared to virtual machine resources. Finally, we introduce a simple im- plementation of real-time serverless to demonstrate its feasibility and capability to support bursty, real-time workloads efficiently.
Original Document
The original document is available in PDF (uploaded 15 July, 2019 by Andrew Chien).