Abstract
Cloud providers offer virtual machines (VM) located in physical machines
(PM) using the “pay-as-you-go” model to satisfy emerging demand for
online computational services. If the instantaneous utilized capacity
requested by VMs exceeds a certain threshold of the total capacity a PM
can offer, a hotspot happens and may cause unacceptable VM performance
degradation. Hotspots can be resolved by relocating some VMs to other
PMs using live migration. However, the problem of selecting which VM(s)
to release is challenging because the utilized capacity demanded by VMs
changes continuously over time. In this work, we propose a Predicted
Mixed Integer Linear Programming (MILP) Robust Solver (PMRS), which
predicts the utilized capacity range of each VM and applies the
Γ-robustness theory to ensure that PM is hotspot-safe with desired
probability. The latter allows us to formulate the hotspot resolution as
a Γ-robust knapsack problem (Γ-RKP) that can be solved by a novel MILP
model. Extensive experiments based on real-trace data and large-scale
synthetic data demonstrate the effectiveness of the PMRS. More
encouragingly, the application of the PMRS in the real-production
environment benefits Huawei Cloud by resolving all existing and 94%+
potential future hotspots with minimal migration overhead.Cloud
providers offer virtual machines (VM) located in physical machines (PM)
using the “pay-as-you-go” model to satisfy emerging demand for online
computational services. If the instantaneous utilized capacity requested
by VMs exceeds a certain threshold of the total capacity a PM can offer,
a hotspot happens and may cause unacceptable VM performance degradation.
Hotspots can be resolved by relocating some VMs to other PMs using live
migration. However, the problem of selecting which VM(s) to release is
challenging because the utilized capacity demanded by VMs changes
continuously over time. In this work, we propose a Predicted Mixed
Integer Linear Programming (MILP) Robust Solver (PMRS), which predicts
the utilized capacity range of each VM and applies the Γ-robustness
theory to ensure that PM is hotspot-safe with desired probability. The
latter allows us to formulate the hotspot resolution as a Γ- robust
knapsack problem (Γ-RKP) that can be solved by a novel MILP model.
Extensive experiments based on real-trace data and large-scale synthetic
data demonstrate the effectiveness of the PMRS. More encouragingly, the
application of the PMRS in the real-production environment benefits
Huawei Cloud by resolving all existing and 94%+ potential future
hotspots with minimal migration overhead.