[storyline] [intro]

Delegating Tiered Memory Management for Virtualized Clouds with HyperTier

Introduction v2

• Cloud faces memory capacity expansion challenges

  • Efficient + elastic + scalable VM is the backbone of today's cloud service
  • memory per CPU trend

• TM briges this gap but requires careful management

• Cloud already presents capacity and performance tradeoff to guests

  • Mem optimized VM instances
  • Blooming OS based TMM support
  • Rethink responsibilities division

• A novel angle with guest-delegated TMM

Introduction (围绕delegation单线讲: 提出delegation->优势->遇到的问题->解决方案)

• VM is at the heart of cloud datacenter

  • Basic service unit
  • Requirements: efficiency + elasticity + scalability

• Industry's goal in using TM and motivation

  • Expand capacity
  • Reduce TCO

• Careful TMM (tracking/classification/migration) is required to leverage TM

  • TM characteristic: fast/small + slow/large
  • frequent accessed -> FMEM

• SOTA are all from host/hypervisor level (shorter)

  • HeteroVisor/OS: trapping to collect guest A-bit info
  • Reuse host-based with MMU notifier backend: collect A-bit from EPT
  • HeMem/Memtis: PEBS in host cannot collect guest accesses
  • insight: A-bits require TLB flushes (->overhead)
  • insight: lack of easy accessible hotness information (->PEBS)

• A novel angle with guest-delegated TMM

  • Host/hypervisor-level mgmt exhibits server mgmt overhead
  • No longer require TLB flushes; utilize easily accessible PEBS hotness
  • How (responsibility)? Leave only provision in host (elasiticity) and enables QoS control
  • Delegate hotness mgmt to guest (light weight mgmt; scalability)

• Challenges: how to ensure efficienty and scalability

  • Tracking: readily available PEBS hotness and eliminate TLB flushes
  • Classification: utilize locality available only in guest
  • Page-exchange: reduce metadata ops (locking; flushes)

• Contribution

  • Proposal of guest-delegation
  • Guest components
  • Provision
  • Evaluation
  • Open source

Background

• (Foundational Knowledge) TMM and its pipeline
• (Prior research) Hypervisor-based TMM
• (Knowledge gap) EPT-friendy PEBS

Motivation

• Hypervisor-based TMM overhead

• E2E principle