Modern cloud applications have increasingly shifted from monolithic architectures to microservices and serverless frameworks, enabling scalable deployment of large-scale services such as social networks, e-commerce, and banking. While these architectures improve modularity and scalability of the system, they also incur significantly higher memory access latency: where a cache miss in a monolithic system typically is resolved locally through the cache hierarchy or DRAM, microservice architectures frequently require remote datastore accesses via RPCs, increasing effective data access latency by several orders of magnitude. However, incoming requests frequently contain sufficient information to predict the data that will be accessed, yet existing datacenter pipelines do not leverage this opportunity. We propose a request-aware pre-fetching architecture that delivers the incoming request to both the scheduler and a prefetching module in parallel. While the scheduler processes the request, the prefetching module retrieves the required data in parallel, reducing memory stalls and improving end-to-end request latency. We will begin by evaluating the request pipelines for several microservice and serverless frameworks, such as for social media requests, to better understand the scope of applications, and we would then simulate typical microservice workloads to establish the opportunity for improvement.

By participating in a SuperUROP, I hope to gain more hands-on experience with a research project from concept to result while learning the process of peer review and paper writing. I also look forward scientific communication and presentation skills, as well as gain feedback from industry experts and faculty.