SUPERFLUIDITY project goals: instantiate network functions and services on-the-fly; run them anywhere in the network (core, aggregation, edge); migrate them transparently to different locations; make them portable across heterogeneous infrastructure environments (computing and networking), while taking advantage of specific hardware features, such as high performance accelerators, when available.
Conclusions: Unikernel virtualization can provide VM instantiation and boot time in the order of ms; ongoing: consolidation of results, generic and automatic optimization process for hypervisor toolstack and for guests. Work is still needed at the level of Virtual Infrastructure Managers e.g. OpenStack (~ 1 s), Nomad (~ 300 ms). VIMs are currently designed for generality, the challenge is to specialize them in a flexible way, keeping the compatibility with the mainstream versions.
Superfluid NFV: VMs and Virtual Infrastructure Managers speed-up for instantaneous service instantiation
1. Superfluid NFV: VMs and Virtual Infrastructure Managers speed-up for
instantaneous service instantiation
Stefano Salsano (CNIT/Univ. of Rome Tor Vergata), Felipe Huici (NEC)
October 10th 2016 – EWSDN @ SDN & OpenFlow World Congress
Joint work with Filipe Manco, Florian Schmidt, Kenichi Yasukata (NEC) - Pier Luigi Ventre,
Claudio Pisa, Giuseppe Siracusano, Paolo Lungaroni, Nicola Blefari-Melazzi (CNIT)
A super-fluid, cloud-native, converged edge system
2. Outline
• The SUPERFLUIDITY project – goals and approach
• Part I – Speed up of:
– Virtualization Platform (including the hypervisor)
– The guests (i.e., virtual machines)
• Part II – Speed up of:
– Virtual Infrastructure Managers
2
3. Outline
• The SUPERFLUIDITY project – goals and approach
• Part I – Speed up of:
– Virtualization Platform (including the hypervisor)
– The guests (i.e., virtual machines)
• Part II – Speed up of:
– Virtual Infrastructure Managers
3
4. SUPERFLUIDITY goals
• Instantiate network functions and services on-the-fly
• Run them anywhere in the network (core, aggregation, edge)
• Migrate them transparently to different locations
• Make them portable across heterogeneous infrastructure environments
(computing and networking), while taking advantage of specific hardware
features, such as high performance accelerators, when available
4
5. SUPERFLUIDITY approach
• Decomposition of network components and services into elementary and reusable
primitives (“Reusable Functional Blocks – RFBs”)
• Native, converged cloud-based architecture
• Virtualization of radio and network processing tasks
• Platform-independent abstractions, permitting reuse of network functions across
heterogeneous hardware platforms
• High performance software optimizations along with leveraging of hardware
accelerators
5
6. SUPERFLUIDITY architecture
6
Based on the on the concept of
Reusable Functional Blocks (RFBs),
applied to different heterogeneous
RFB Execution Environments (REE)
Different RDCLs (RFB Description and
Composition Languages) can be used in
different environments.
7. • Classical NFV environments (i.e. by ETSI NFV standards)
– VNFs are composed/orchestrated to realize Network Services
– VNFs can be decomposed in VNFC (VNF Components)
«Big»
VNF
«Big»
VNF
«Big»
VNF
«Big»
VNF
VNF
C
VNF
C
VNF
C
VM
VM
VM
Heterogeneous composition/execution environments
7
8. Heterogeneous composition/execution environments
• Towards more «fine-grained» decomposition…
• Modular software routers (e.g. Click)
– Click elements are combined in configurations (Direct Acyclic Graphs)
8
9. Heterogeneous composition/execution environments
• Towards more «fine-grained» decomposition…
• XSFM-based (eXtended Finite State Machine) decomposition of traffic forwarding /
flow processing tasks, and HW support for wire speed execution
9
10. Network Functions reuse/composition
NFV-like VNF
management
General purpose
Computing Platform (CPUs)
specific
VNF
VM
specific
VNF
VM
SDN-like
Configuration
deployment
The ‘traditional’ VNF’s view
General purpose computing platform
Full flexibility (VNF = ‘anything’ coded in ‘any’ language)
Performance limitations (slow path execution)
Pre-implemented
match/action table
OpenFlow
(HW) switch
Flow table Entry
Flow table Entry
Flow table Entry
flow-mod
Traditional SDN southbound (OpenFlow)
Domain-specific platform (OpenFlow router)
Extremely limited flexibility (hardly an NF)
Line-rate performance (TCAM/HW)
10
11. General purpose
Computing Platform (CPUs)
specific
VNF
VM
specific
VNF
VM
The ‘traditional’ VNF’s view
General purpose computing platform
Full flexibility (VNF = ‘anything’ coded in ‘any’ language)
Performance limitations (slow path execution)
Pre-implemented
match/action table
OpenFlow
(HW) switch
Flow table Entry
Flow table Entry
Flow table Entry
flow-mod
Traditional SDN southbound (OpenFlow)
Domain-specific platform (OpenFlow router)
Extremely limited flexibility (hardly an NF)
Line-rate performance (TCAM/HW)
NFV-like VNF
management
SDN-like
Configuration
deployment
Lean towards ‘more
domain specific’
network computing
HW
Lean towards ‘more
expressive’ programming
constructs / APIs
Network Functions reuse/composition
11
12. APIs definition
RFB
#a
RFB
#b
RFB
#c
RFB
#n
REE - RFB Execution Environment
(node-level) RDCL script
REEREE
RFB#2 RFB#3
(network-wide) REE - RFB Execution Environment
(network-level) RDCL script
RFB#1
REE
Manager
REE User
REE
Resource
Entity
UM API
MR API
REE User
REE
Manager
UM API
REE
Resource
Entity
MR API
RDCLs (RFB Description and
Composition Languages) are used
on the logical API between the
“user” of an RFB Execution
Environment and the “manager”
(provider) of such environment
Different RDCLs can be used in
different environments.
12
13. Rationale for the unified RFB concept
• It is not a top-down approach: we cannot impose a single model and apply it in all
environments
• Convergence across different heterogeneous environments (where possible)
– Unify/combine the languages and tools
• Helps to identify how the different environments can share resources and can be
combined in a common infrastructure
13
14. Convergence approach
A unified cloud platform for radio and network functions. CRAN, MEC and cloud technologies are
integrated with an architectural paradigm that can unify heterogeneous equipment and
processing into one dynamically optimised, superfluid, network 14
15. Towards sub 10 ms service instantiation
• The SUPERFLUIDITY project – goals and approach
• Part I – Speed up of:
– Virtualization Platform (including the hypervisor)
– The guests (i.e., virtual machines)
• Part II – Speed up of:
– Virtual Infrastructure Managers
15
16. Why a superfluid NFV (sub 10 ms service instantiation)
• Quick provisioning of services: JIT proxies, firewalls, on-the-fly monitoring
• Quick migration of services: base station splitting
• Optimized use of resources thanks to dynamic sharing
• Hosting large number of services on the same server: e.g., vCPE
• High-performance networking: NFV, virtualized CDNs, etc.
• Quick-checkpointing
• General investment and operating cost reductions
16
19. VM instantiation and boot time
19
Orchestrator
request
VIM
operations
Virtualization
Platform
Guest OS (VM)
Boot time
1-2 s
5-10 s
~1 s
20. Towards sub 10 ms service instantiation
• The SUPERFLUIDITY project – goals and approach
• Part I – Speed up of:
– Virtualization Platform (including the hypervisor)
– The guests (i.e., virtual machines)
• Part II – Speed up of:
– Virtual Infrastructure Managers
20
22. But I need to pick my poison ☹
Lightweight
Iffy isolation
CONTAINERS HYPERVISORS
Strong isolation
Heavy weight
We need a superfluid virtualization
22
24. Towards a Superfluid Platform
• Fast boot/destroy/migration times
• Reducing guest memory footprints
• Optimizing packet I/O (40-80 Gb/s)
• New hypervisor schedulers
24
25. Towards a Superfluid Platform
• Fast boot/destroy/migration times
• Reducing guest memory footprints
• Optimizing packet I/O (40-80 Gb/s)
• New hypervisor schedulers
25
26. A Quick Xen Primer
Dom0 (Linux/NetBSD)
Hardware (CPU, Memory, MMU, NICs, …)
Xen Hypervisor
libxc libxs
libxl toolstack
xl
NIC
drivers
block
SW switch
virt
drivers
netback
xenbus
DomU 1
netfront
xenbus
OS (Linux)
apps
Xen
store
26
27. A Unikernel Primer
• Specialized VM: single
application + minimalistic OS
• Single address space,
co-operative scheduler so low
overheads
driver1
driver2
app1
GENERAL-PURPOSE
OPERATING SYSTEM
(e.g., Linux, FreeBSD)
KERNELSPACEUSERSPACE
app2
appNdriverN
Vdriver1
vdriver2
app
MINIMALISTIC
OPERATING SYSTEM
(e.g., MiniOS, OSv)
SINGLEADDRESS
SPACE
27
28. Memory Footprint
• Xen allocates a minimum of 4MB for all guests, irrespective
of how much memory is needed or asked for
– Modified the toolstack to allow memory allocations to be
specified in KBs
• Guests require a lot of memory to run
– Use unikernels instead
28
29. Memory Footprint - Result
• Hello world guest
– 296KB
• Ponger guest 692KB
– 350KB come from lwip and newlibc
• This is with minor optimizations to MiniOS
(e.g., reducing the threads’ stack size)
29
30. VM Boot Times
1. xl create myvm.cfg
2. libxl (e.g., parse config)
3. libxc (e.g., hypercalls to create guest, reserve memory, load image into memory)
4. Write entries to Xenstore for guest to use
5. Boot guest
6. Guest retrieves information from Xenstore (e.g., even channels, back-end
domains)
Note: VM destroy and migration times depend on
similar toolstack/Xenstore operations!
30
31. Main Culprits
• Toolstack
– Inefficient/outdated code
– Too generic for our purposes (e.g., support for HVM guests, QEMU).
• Xenstore
– Used to communicate information between guests (e.g., event channel
numbers, back-end domain information)
– Relies on transactions, watches
– Single point of failure, bottleneck
• And of course the guest
– Use unikernels
31
32. Towards a Solution
• Toolstack – Chaos
– Complete re-write of toolstack, no need for libxl/libxc
– Includes framework for easily plugging in different elements of a toolstack
(e.g., with or without Xenstore)
• Xenstore
– Do we really need one?
– Design and implementation of “Xenstore-less” guests and the corresponding
toolstack
32
40. Virtualization Platforms & Guests - Ongoing & Future Work
• Short term
– Lots of clean-up, more results
– Libxc replacement
– High performance (40-80 Gb/s) service chaining
• Longer term
– New hypervisor schedulers for massive consolidation, high packet I/O
– Unicore: tools for automatically building high performance unikernels
and OSes → OS-level decomposition
40
41. Towards sub 10 ms service instantiation
• The SUPERFLUIDITY project – goals and approach
• Part I – Speed up of:
– Virtualization Platform (including the hypervisor)
– The guests (i.e., virtual machines)
• Part II – Speed up of:
– Virtual Infrastructure Managers
41
42. VM instantiation and boot time
42
Orchestrator
request
VIM
operations
Virtualization
Platform
Guest OS (VM)
Boot time
1-2 s
~1 ms
~1 ms
• Unikernels can provide low
latency instantiation times for
“Micro-VNF”
• What about VIMs (Virtual
Infrastructure Managers) ?
43. Performance analysis and Tuning of VIMs for Micro VNFs
• General model of the VNF instantiation process
• Modifications to VIMs to instantiate Micro-VNFs based on
ClickOS Unikernel
• Methodology to evaluate the performances
• Performance Evaluation
43
44. Virtual Infrastructure Managers (VIMs)
We considered the performance of two VIMs :
• OpenStack Nova
– OpenStack is composed by subprojects
– Nova: orchestration and management of computing resources ---> VIM
– 1 Nova node (scheduling) + several compute nodes (which interact with the hypervisor)
– Not tied to a specific virtualization technology
• Nomad by HashiCorp
– Minimalistic cluster manager and job scheduler
– Nomad server (scheduling) + Nomad clients (interact with the hypervisor)
– Not tied to a specific virtualization technology
44
49. VIM modifications to instantiate (ClickOS) Micro VNFs
49
A regular VM can boot its OS
from an image or a disk snapshot
that can be read from an
associated block device (disk).
The host hypervisor instructs the
VM to run the boot loader, which
reads the kernel image from the
block device.
ClickOS based MicroVNFs, are
shipped as a tiny kernel without
a block device. These VMs need
to boot from a so-called diskless
image. The host hypervisor reads
the kernel image from a file or a
repository and directly injects it
in the VM memory.
Virtual
Infrastructure
Manager
Virtualization
Platform
(Hypervisor)
This interface needs to
be modified to support
the boot of “diskless
images”
50. VIM modifications to instantiate (ClickOS) Micro VNFs
• OpenStack
– Xen supported out of the box, using the Libvirt toolstack
– We considered the boot of diskless images targeting only one component
(Nova Compute) and a specific toolstack, Libvirt.
– Libvirt talks with Xen using libxl the default Xen toolstack API.
– We modified the XML description of the guest domain provided by the driver,
changing the XML description on the fly before the creation of the domain.
• Nomad
– Xen not supported out of the box
– We developed a new Nomad driver for Xen, called XenDriver .
– The new driver communicates with the XL Xen toolstack and it is also able to
instantiate a ClickOS VM.
50
51. VIM performance evaluation approach
• We evaluate the VM scheduling and instantiation phase, combining message trace
analysis and timestamps in the code
• Message traces (coarse information, beginning and end of the different phases)
– VIM Message Analyzer capable of analyzing Nova and Nomad message exchanges
• Detailed breakdown with timestamps in the code (Nomad Client, Nova Compute)
• Workload generators:
– OpenStack : Rally benchmarking tool
– Nomad : developed the “Nomad Pusher”, a utility written in the GO language which
programmatically submits jobs to the Nomad Server.
51
52. Results – ClickOS instantiation times
52
OpenStack Nova
Nomad
seconds
seconds
53. There is no comparison implied…
• NB: the purpose of the work is NOT to compare OpenStack vs. Nomad.
The goal is to understand how both behave and find ways to reduce
instantiation times.
• A direct comparison makes few sense. OpenStack is a much more
complete framework in terms of offered functionality and different
types of supported hypervisors. Moreover, the comparison is unfair
also because for the Nomad case we have developed a driver only
targeted to support the Xen/Click OS case.
53
54. VIM Tuning
• OpenStack
– Diskless VM -> we can skip most of the actions performed during the image creation;
– UniKernels are special purpose VMs:
• SSH is really needed ?
• Full-IP stack ?
– We were able to reduce the spawning time of about 70%
– Looking at the overall instantiation time, the relative reduction is about 45%;
• Nomad
– No much space for the optimization;
• We implemented only the necessary functionality;
– We introduced further improvements assuming a local store for the Micro VNFs,
reducing the Driver operation of about 30 ms;
54
57. VIM performances - Ongoing & Future Work
• Consider the impact of system load on the performance
– Measure the average instantiation times considering batches of incoming requests with given
rate (requests/s) and arrival patterns.
– Analyze the impact of the number of already allocated VMs and of the number of target nodes
to be deployed.
• Keep improving the performance of the considered VIMs
– e.g. trying to replace the lazy notification mechanism of Nomad with a reactive approach
• Extend the analysis to another VIM
– OpenVIM from the OSM project
57
58. Unikernel virtualization in the SUPERFLUIDITY vision
• We have considered the optimization of Unikernel virtualization and the needed
enhancements to Virtual Infrastructure Managers to support Unikernels.
• In the SUPERFLUIDITY vision, Unikernels are interesting as they support the
decomposition of network services in “smaller” components that can be deployed
on the fly.
• The NFV Infrastructure should be extended in order to support Unikernel
virtualization in addition to traditional VMs. This way it will be possible to design
services that exploit the most efficient solutions depending on several factors.
58
59. Conclusions
• Unikernel virtualization can provide VM instantiation and boot time in
the order of ms
– ongoing: consolidation of results, generic and automatic optimization process for
hypervisor toolstack and for guests
• Work is still needed at the level of Virtual Infrastructure Managers
– e.g. OpenStack (~ 1 s), Nomad (~ 300 ms)
• VIMs are currently designed for generality, the challenge is to specialize
them in a flexible way, keeping the compatibility with the mainstream
versions
59
60. References - SUPERFLUIDITY
• SUPERFLUIDITY project Home Page http://superfluidity.eu/
• G. Bianchi, et al. “Superfluidity: a flexible functional architecture for 5G
networks”, Transactions on Emerging Telecommunications Technologies
27, no. 9, Sep 2016
60
61. References – Speed up of Virtualization Platforms / Guests
• J. Martins, M. Ahmed, C. Raiciu, V. Olteanu, M. Honda, R. Bifulco, F. Huici,
“ClickOS and the art of network function virtualization”, NSDI 2014, 11th
USENIX Conference on Networked Systems Design and Implementation,
2014.
• F. Manco, J. Martins, K. Yasukata, J. Mendes, S. Kuenzer, F. Huici,
“The Case for the Superfluid Cloud”, 7th USENIX Workshop on Hot Topics
in Cloud Computing (HotCloud 15), 2015
61
62. References – Speed up of VIMs
• P. L. Ventre, C. Pisa, S. Salsano, G. Siracusano, F. Schmidt, P. Lungaroni, N.
Blefari-Melazzi,
“Performance Evaluation and Tuning of Virtual Infrastructure Managers
for (Micro) Virtual Network Functions”, IEEE NFV-SDN 2016 Conference,
Palo Alto, USA, 7-11 Nov. 2016
62
63. Thank you. Questions?
Contacts
SUPERFLUIDITY project, Speed up of VIMs
Stefano Salsano, Associate Professor
University of Rome Tor Vergata / CNIT
stefano.salsano@uniroma2.it
Speed up of Virtualization Platforms / Guests
Felipe Huici, Chief Researcher
Networked Systems and Data Analytics Group
NEC Laboratories Europe
felipe.huici@neclab.eu
63
64. The SUPERFLUIDITY project has received funding from the European Union’s Horizon
2020 research and innovation programme under grant agreement No.671566
(Research and Innovation Action).
The information given is the author’s view and does not necessarily represent the view
of the European Commission (EC). No liability is accepted for any use that may be
made of the information contained.
64