Abstract
Memory overcommitment is a fundamental technique in cloud computing that enables providers to maximize resource utilization by allocating more memory to virtual machines than physically available on the host. However, this practice introduces a critical performance anomaly known as double swapping, where both the host and guest independently swap the same memory pages to disk, resulting in redundant I/O operations.
This paper introduces ZeroSwap, a lightweight coordination mechanism that eliminates double swapping.
ZeroSwap intercepts guest I/O operations and distinguishes between normal disk writes, which must be preserved for data consistency, and swap-related writes that are candidates for optimization.
ZeroSwap requires no modifications to guest operating systems and imposes negligible overhead.
We evaluated ZeroSwap on modern datacenter hardware configurations, including high-speed SSDs and disaggregated storage architectures, using workloads from NPB, SPEC CPU2017, and Memcached. Our results demonstrate that ZeroSwap achieves application speedups of up to 7.06x.