Graduate Thesis Or Dissertation

 

Fluidmem: Open Source Full Memory Disaggregation Public Deposited

https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/qn59q4253
Abstract
  • To satisfy the performance demands of memory-intensive applications facing DRAM shortages, the focus of previous work has been on incorporating remote memory to expand capacity. However, the emergence of resource balancing as a priority for cloud computing requires the capability to dynamically size virtual machine memory up and down. Furthermore, hardware-based or kernel space implementations hamper flexibility with respect to making customizations or integrating the continuing open source advancements in software infrastructure for the datacenter. This thesis presents an architecture to meet the performance, bi-directional sizing, and flexibility challenges of memory disaggregation in the cloud. The implementation, called FluidMem, is open source software that integrates with the Linux kernel, KVM hypervisor, and multiple key-values stores. With FluidMem, a virtual machine's local memory can be transparently extended or entirely transferred to a remote key-value store. By fully implementing the dynamic aspect of data center memory disaggregation, FluidMem allows a VM's footprint to be precisely sized, expandable for application demands, but it leaves cloud operators with a non-intrusive recourse if memory becomes scarce. Full memory disaggregation in FluidMem enables the local memory footprint of a Linux VM to be scaled down to 180 pages (720 KB), yet still accept SSH logins. A user space page fault handler in FluidMem uses the userfaultfd mechanism in the Linux kernel to relocate any page of the VM to remote memory and outperforms swap-based paging. Page fault latencies via FluidMem to RAMCloud are 40% faster than the RDMA remote memory swap device in the Linux kernel, and 77% faster than SSD swap. FluidMem's remote memory expansion performance with three key-value backends is evaluated against swap-based alternatives for a MongoDB workload and the Graph500 benchmark.
Creator
Date Issued
  • 2019
Academic Affiliation
Advisor
Committee Member
Degree Grantor
Commencement Year
Subject
Last Modified
  • 2019-11-14
Resource Type
Rights Statement
Language

Relationships

Items