Aquifer: New memory pooling cuts MicroVM cold starts 2.2x
Wasted DRAM costing 25-35% in cloud clusters gets fixed by hybrid CXL+RDMA.
Aquifer, developed by Junliang Hu, Huaicheng Li, and Ming-Chang Yang, addresses memory stranding — which wastes 25-35% of installed DRAM in production cloud clusters. The system uses a hierarchical memory pool combining CXL (low-latency, load/store-transparent but limited to a pod) and RDMA (cluster-wide reach at higher latency with software overhead). By analyzing MicroVM snapshot images, they found most pages are zero or cold, enabling a hotness-based format that eliminates zero pages and places only the hot working set in CXL memory while cold pages go to RDMA. Aquifer also introduces an ownership-based coherence protocol for CXL 2.0 multi-headed devices lacking hardware cache coherence.
Aquifer employs a copy-based page serving mechanism: before MicroVM resume, it pre-installs hot pages from CXL memory, and asynchronously demand-pages cold pages from RDMA. On emulated CXL+RDMA hardware, Aquifer achieves a 2.2x geometric-mean speedup in end-to-end invocation time over Firecracker (the standard MicroVM) and 1.1x over the next best alternative. This represents the first practical exploration of hierarchical memory pooling for serverless computing, directly targeting the cold-start latency bottleneck.
- Memory stranding wastes 25-35% of DRAM in production clouds — Aquifer repurposes that via CXL+RDMA.
- Achieves 2.2x speedup over Firecracker and 1.1x over the next best alternative on emulated hardware.
- Hotness-based snapshot format eliminates zero pages and places hot pages in CXL, cold in RDMA for optimal latency.
Why It Matters
This could slash cold-start delays in serverless platforms, saving cloud providers millions in DRAM costs.