SnowFlock:Rapid Virtural Machine Cloning for Cloud Computing

1
SnowFlockRapid Virtural Machine Cloning for
Cloud Computing

• H. Andrés Lagar-Cavilla, Joseph A. Whitney, Adin
  Scannell, Philip Patchin, Stephen M. Rumble,Eyal
  de Lara, Michael Brudno, Satyanarayanan
• University of Toronto and Carnegie Mellon
  University

2
Outline

• Background
• Design and building block of the P2P_VoD system
• Performance metrics and measurement methodology
• Measurement result and analysis
• Future work

3
Background

• Cloud computing shifts the hardware and staffing
  costs of managing a computational center to third
  parties.
• Major advantage of cloud computing is the
  ability to use a variable number of physical
  machines and VM instances depending on the needs
  of the problem
• Shortage
• Instantiating new VMs is a slow operation
• Unaware of the current application state
• Idle VMs are likely to be consolidated and
  swapped out, incurring costly migration delays
  before they can be used

4
Related work

• Area of replication
• Potemkin(Vrable 2005)short-lived lightweight VMs
  cloned from a static template in the same machine
  with memory copy-on-write techniques
• Remus(Cully 2008)instantaneous failover by
  keeping an up-to-date replica of a VM in a
  separate host
• Copy on reference
• Theimer 1985, Zayas 1987 precursor to our
  memory-on-demand technique
• Lagar-Cavilla 2007,Sapuntzakis 2002,Kozuch 2002
  lazy copy-on reference for VM disk state
• Capability
• Amazon EC2follows industry standard techniques
  for the provisioning of VMs on the
  fly(unpublished)consolidtion via memory sharing
  live migration ballooning,resuming from
  disk(Steinder 2007,Waldspurger 2002, VMotiion
  2005)

5
Related work

• SnowFlock take advantage of
• Emeneker 2007,Foster 2006 ,Chase 2003virtual
  cluster focusing on resource provisioning and
  management
• Usher(McNett 2007)a manager of clusters of VMs
  that could be plugged in as a SnowFlock resource
  manager
• Emulab(Hibler 2008)instantiate dozens of nodes
  for a network emulation experiment
• Frisbee(hibler 2003) multicast distribution to
  apply disk images to nodes during experiment
  setup
• RDMA and Infiniband(huang 2007)high-speed
  interconnects

6
VM fork

• familiar to process fork, each child have a vmid
• Include integration with a dedicated, isolated
  virtual network connecting child VMs with their
  parent
• Usage model the ephemeral nature of children
• Replicate all the processes and threads of the
  parent VM

7
VM fork differences

• Forked copy can instantiated on a set of
  different physical machines that takes advantage
  of large compute clusters.Compare to Vrable
  2005
• Parallel enabling the creation of multiple child
  VMs with a single call
• Replicates all of the processes and threads of
  the originating VM

8
VM fork
Several useful and well-known patterns that are
based on stateful replication
9
Design Rationale

• Network latency

10
Four insights in implementation

• Possible to start executing a child VM on a
  remote site by initially replicating only minimal
  state
• Children will typically access only a fraction of
  the original memory image of the parent
• Be common for children to allocate memory after
  forking
• Children often execute similar code and access
  common data structures

11
SnowFlock implementation

• Implemented as a combination of modifications to
  the Xen VMM and daemons that run in d0
• Based on lazy state replication combined with
  avoidance heuistics to minimize state transfer
• Steps of SnowFlock

12
API
13
VM Descriptors

• Metadata describing the VM and its virtual
  devices
• A few memory pages shared between the VM and the
  Xen hypervisor
• The registers of the main VCPU
• GDT used by the x86 segmentation hardware for
  memory protection
• The page tables of VM

14
Replication descripter Time
15
Memory-On-Demand

• Memtap impliments A copy-on-access policy for
  the clone VMs memory
• Use shadow page tables
• parent VM implements a copy-on-write policy to
  serve the memory image
• SMP-safe shared bitmap is used by Xen and memtap
  to the presence of the memory

16
Memtap,Heuristics,multicast
17
Avoidance Heuristics

• To bypass numbers of unnecesary memory fetches
  while retaining correctness, two fetch avoidance
  policies are provided
• Optimize the general case in which a clone VM
  allocates new state
• Addresses the case where a virtual IO device
  writes to the guest memory

18
Multicast Distribution

• Switch programming
• use IP-multicast in order to send data to
  multiplehosts simultaneously
• Receive pages asynchronously and unpredictably in
  response to requests by fellow VM clones
• Flow control logic
• using a weighted average of the number of bytes
  transmitted or received every ten ms
• lockstep detection, which aims to leverage the
  similarity in memory access patterns across clones

19
Virtural IO Devices in SnowFlock

• SnowFlock provides a Virtual Disk for each clone
  VM
• COW slice rendering the previous state of the
  disk immutable
• For data-intensive tasks, use NFS, Hadoop or
  Lustre
• Guarantees secure network isolation
• Package level rewritting
• ARP

20
Evaluation

• 3 typical applications from bioinformatics and 3
  applications representative of the fields of
  graphics rendering, parallel compilation, and
  financial services.
• They are NCBI BLAST, SHRiMP, ClustalW, QuantLib,
  Aqsis-Renderman,Distcc

21
Comparison
22
Single thread zero-cost
23
Cycle cloning
24
Future Work

• Interactions of VM fork with data parallel APIs
  such as MapReduce
• Push VM state in the background
• Apply SnowFlock techniques to wide-area VM
  migration

25
Thanks