Lecture 31

Resource Management in VMM

Shares vs Working sets

• OS me working set model hota, kya pages abhi being used
• in cloud, we have to also see Vm level promises we make
  • importance, ownership, amount of money paid to service provider for executing the vm
  • penalize less imp bm, even if it would derive largest perf benefit from additional memory

Share based allocation

• give resources acc to shares that is held by vm
• share is regulated parameter
• client consumes resources proportional to its share

• guarantted min resource fraction equal to fraction of its total shares in system

• algorithms for proportional share
  • dynamic min-funding revocation algo
  • revoke memory from client that owns fewest shares per allocated page
  • take memory away from those payingh lower price to those willing to pay higher price

VMWare ESX - Share based allocation

• always not feasible
• might be ki higher ratio wala having larger #pages, ab woh use karra ki nahi idk

• idle memory tax

• prob with pure proportional share allocn
  • idle VMs with many shares can hoard memory unproductively
  • active vms with few shares suffer under memory pressure
• resolve this by introducing idle memory tax
  • fraction of pages being utilized and idle nikal le
  • f = fraction of number of pages which r active (say 10 me se 5 hi active => 0.5)
  • cost is k = (1/(1-t)), t = tax
  • if t = 0, same as min-funding algo = S/P
  • t defaults to 0.75
    • cost we r imposing on system with idle pages
  • penalizing not utilized pages

• jiske pas extra tha usse le lia, and jiske pas nahi tha, usko de dia

Resouce allocation policies

• 3 parameters
  • min
    • min this should be avaialnle so it can run
  • max
    • amount of physical memory configured for use by guest os
  • shares
    • entitle vm to fraction of physical memory based on proportional share allocation policy
• Admission control
  • a vm is admitted if min+overhead(for page tables, shadow page tables + grpahics grame buffer) is available
  • max-min swap space honi chahiye
  • tabhi take this vm into ur system
• ESX dynamically computes memory allocations and reclaims memory when below threshold
  • 4 memory thresholds
    • high
      • no reclamation
      • agar free memory available is only 6%, then lio
    • soft
      • use ballooning to reclaim
      • free% = 4% pe
    • hard
      • forcible paging
      • 2%
    • low
      • stop execution of vms reclaim pages
      • 1%

Sharing MEmory across VMs

• if 2 processes using same code, say vi editor, os only creates one copy and maps to the process executing
• esx bhi aisa kuch karta
• say 3 vms hai, 2 runnign windows, 1 running linux
• if 2 vms use same pages.. how does it know?
  • hashing technique use karle
  • and idea laga ki agar same pages acces kar rahe 2 machine
  • if changes kie kisi ne, then copy karde(copy on write policy)

Resource mgmt - local vs global

• optimizing reosource mmgmt at vmm level(locally)
• but it might not globally lead to optimal resource mgmt
• cloud me globally chahiye bhai
• 2 parameters that paly role in dynamic resource mgmt
  • perf
    • energy efficiency
  • vm migration
• why imp to look globally?
  • even if har vm khud se acha karri, it might be underloaded, vms might be able to consolidated

Borg system

• cluster manager used by google for admitting, scheduling, starting, restarting and monitoring jobs

Architecture

• borg cell
  • set of machines
  • a borgmaster
    • logically centralized controller
  • an agent process = borglet
    • runs on each machine
    • gives info to server

• machines in a cell belong to single cluster
  • cluster usually hosts one large cell and may have smaller scale test or special purpose cells

Borgmaster

• schedules jobs submitted to it
• logically a single process but is actually replicaated 5 times
• uses paxos to manage shared state

Borglet

• software running on each machine
• starts and stops a job

Workload Characteristics

• long running services
  • should nvr go down
  • very less latency
• batch jobs

• production and batch jobs
  • production = latency sensitive
  • batch = lower priority

Clusters and cells

• user submits jobs to borg mgr
• baki borg master dekhta
• kaha run karu
• kitne resources du
• kya dependencies hai, install kar usko
• monitor machines if fail, etc

Jobs and Tasks

Priority, quota and admission control

• every job has a priority
• high priority can preempt lower priority for resources
• preemption cascades ho sakte
  • solution: same band wale nahi preempt kar sakte
• ands
  • monitoring>production>batch>best effort(aka testing/free)

Naming and Monitoring

Scheduling