Internet Research Task Force (IRTF) R. Krishnan
Internet Draft N. Figueira
Category: Experimental Brocade
Dilip Krishnaswamy
IBM Research
D. R. Lopez
Telefonica I+D
Steven Wright
AT&T
Tim Hinrichs
VMware
Expires: May 2015 November 25, 2014
NFVIaaS Architectural Framework for Policy Based Resource Placement
and Scheduling
draft-krishnan-nfvrg-policy-based-rm-nfviaas-03
Abstract
One of the goals of Network Functions Virtualization (NFV) is to
offer the NFV infrastructure as a service to other SP customers -
this is called NFVIaaS. Virtual Network Function (VNF) deployment in
this paradigm will drive support for unique placement policies,
given VNF's stringent service level specifications (SLS) required by
customer SPs. Additionally, NFV DCs often have capacity, energy and
other constraints - thus, optimizing the overall resource usage
based on policy is an important part of the overall solution. The
purpose of this document is to depict an architectural framework for
policy based resource placement and scheduling in the context of
NFVIaaS.
Status of this Memo
This Internet-Draft is submitted to IETF in full conformance with
the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as Internet-
Drafts.
Internet-Drafts are draft documents valid for a maximum of six
months and may be updated, replaced, or obsoleted by other documents
Krishnan Expires May 2015 [Page 1]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
at any time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt.
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.
This Internet-Draft will expire in May 2015.
Copyright Notice
Copyright (c) 2014 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with
respect to this document.
Conventions used in this document
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119.
Table of Contents
1. Introduction...................................................3
2. NFVIaaS Architectural Framework for Policy Based Resource
Placement and Scheduling..........................................3
3. System Analysis in a OpenStack Framework.......................5
4. Related Work...................................................8
5. Summary........................................................8
6. Future Work....................................................8
7. IANA Considerations............................................8
8. Security Considerations........................................8
9. Contributors...................................................8
10. Acknowledgements..............................................9
11. References....................................................9
11.1. Normative References.....................................9
11.2. Informative References...................................9
Authors' Addresses...............................................10
Krishnan Expires May 2015 [Page 2]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
1. Introduction
One of the goals of NFV [ETSI-NFV-WHITE] is to offer the NFV
infrastructure as a service to other SP customers - this is called
NFVIaaS [ETSI-NFV-USE-CASES]. In this context, it may be desirable
for a Service Provider to run virtual network elements (e.g.,
virtual routers, virtual firewalls, and etc. - these are called
Virtual Network Functions - VNF) as virtual machine instances inside
the infrastructure of another Service Provider. In this document, we
call the former a customer SP and the latter an NFVIaaS SP.
There are many reasons for a customer SP to require the services of
an NFVIaaS SP, including: to meet performance requirements (e.g.,
latency or throughput) in locations where the customer SP does not
have physical data center presence, to allow for expanded customer
reach, regulatory requirements, and etc.
As VNFs are virtual machines, their deployment in such NFVIaaS SPs
would share some of the same placement restrictions (i.e., placement
policies) as those intended for Cloud Services. However, VNF
deployment will drive support for unique placement policies, given
VNF's stringent service level specifications (SLS) required/imposed
by customer SPs. Additionally, NFV DCs often have capacity, energy
and other constraints - thus, optimizing the overall resource usage
based on policy is an important part of the overall solution.
The purpose of this document is to depict an architectural framework
for policy based resource placement and scheduling in the context of
NFVIaaS.
2. NFVIaaS Architectural Framework for Policy Based Resource Placement
and Scheduling
The policy engine performs policy-based resource placement and
scheduling of Virtual Machines (VMs) in support for NFVIaaS. It
determines optimized placement and scheduling choices based on the
constraints specified in the policy. The NFVIaaS Architectural
Framework for Policy Based resource placement and scheduling is
based on the NFV policy architectural framework [IRTF-NFV-POLICY-
ARCH].This is depicted in Figure 1.
In one instantiation of this architecture, the policy engine would
interface with the Measurement Collector to periodically retrieve
instantaneous per-server CPU utilization, which it would then use to
compute a table of per-server average CPU utilization. In an
alternative instantiation of this architecture, the measurement
collector could itself compute per-server average CPU utilization.
Krishnan Expires May 2015 [Page 3]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
The latter approach reduces overhead, since it avoids too frequent
pulling of stats from Ceilometer. The policy engine evaluates such
policies based on an event trigger or a based on a programmable
timer.
Other average utilization parameters such as VM CPU utilization, VM
Memory utilization, VM disk read IOPS etc. could be used by the
policy engine to enforce other types of placement policies.
|---------------------------------------|
| Policy Engine |
|Performs resource placement |
|and scheduling function (proactive and |
|dynamic policy enforcement) |
|--------------------|------------------|
|
|
|-----------|---------------|
| Measurement Collector |
| VM DB - CPU |
| Utilization etc. |
| |
|---------------------------|
Figure 1: NFVIaaS Architecture for Policy Based Resource Placement
and Scheduling
In an ETSI NFV Architectural Framework [ETSI-NFV-ARCH][NFV-MANO-
SPEC], Policy Engine is part of the Orchestrator and the Measurement
Collector is part of the Virtual Infrastructure Manager (VIM).
Krishnan Expires May 2015 [Page 4]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
3. System Analysis in a OpenStack Framework
Consider an NFVIaaS SP that owns a multitude of mini NFV data
centers managed by OpenStack [OPENSTACK] where,
- The Policy Engine function is performed by OpenStack Congress
[OPENSTACK-CONGRESS-POLICY-ENGINE]
- The Measurement Collector function is performed by OpenStack
Celiometer [OPENSTACK-CELIOMETER-MEASUREMENT]
Exemplary Mini NFV DC configuration:
An exemplary mini NFV DC configuration is depicted below.
- There are 210 physical servers in 2U rack server configuration
spread over 10 racks.
- There are 4 types of physical servers each with a different
system configuration and from a particular manufacturer. It is
possible that the servers are from the same or different
manufacturer. For the purpose of this example, server type 1 is
described further. Server type 1 has 32 virtual CPUs and 128GB DRAM
from manufacturer x. Assume 55 physical servers of type 1 per mini
NFV DC.
- There are 2 types of instances large.2 and large.3, which are
described in the table below. Each parameter has a minimum guarantee
and a maximum usage limit.
|--------|----------------- |-------------------|---------------|
|Instance|Virtual CPU Units |Memory (GB) |Minimum/Maximum|
|Type |Minimum Guarantee |Minimum Guarantee |Physical Server|
| |/Maximum Usage |/Maximum Usage |Utilization (%)|
|--------|------------------|-------------------|---------------|
|large.2 | 0/4 | 0/16 | 0/12.5 |
|large.3 | 0/8 | 0/32 | 0/25 |
|--------|------------------|-------------------|---------------|
Table 1: NFVIaaS Instance Types
Krishnan Expires May 2015 [Page 5]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
For the purpose of this example, the Mini NFV DC topology is
considered static -- the above topology, including the network
interconnection, is available through a simple file-based interface.
Policy 1 (an exemplary NFV policy):
Policy 1 is an exemplary NFV policy. In a descriptive language,
Policy 1 is as follows - "For physical servers of type 1, there can
be at most only one physical server with average overall utilization
less than 50%."
The goal of this policy is to address the energy efficiency
requirements described in the ETSI NFV Virtualization Requirements
[ETSI-NFV-REQ].
Various Module Interactions for Policy 1:
The various module interactions with respect the architectural
framework in Figure 1 for Policy 1 is described below.
The policy calls for the identification of servers by type.
OpenStack Congress would need to support server type, average CPU
utilization, and be able to support additional performance
parameters (in the future) to support additional types of placement
policies. OpenStack Congress would run the policy periodically or
based on events such as deleting/adding VMs etc. Initially, we could
use a periodic timer based approach. In case OpenStack Congress
detects a violation, it determines optimized placement and
scheduling choices so that the policy is not violated.
OpenStack Congress could interface with OpenStack Celiometer to
periodically retrieve instantaneous per-server CPU utilization,
which it would then use to compute a table of per-server average CPU
utilization. Alternatively, OpenStack Celiometer could itself
compute per-server average CPU utilization which could be used by
OpenStack Congress.
The proposed module interactions in this NFVIaaS placement policy
example are as depicted in the architectural framework in Figure 1.
A key goal of Policy 1 above is to ensure that servers are not kept
under low utilization, since servers have a non-linear power profile
and exhibit relatively higher power wastage at lower utilization.
For example, in the active idle state as much as 30% of peak power
is consumed. At the physical server level, instantaneous energy
consumption can be accurately measured through IPMI standard. At a
customer instance level, instantaneous energy consumption can be
Krishnan Expires May 2015 [Page 6]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
approximately measured using an overall utilization metric, which is
a combination of CPU utilization, memory usage, I/O usage, and
network usage. Hence, the policy is written in terms of overall
utilization and not power usage.
For an exemplary maximum usage scenario, 53 physical servers could
be under peak utilization (100%), 1 server (server-a) could be under
partial utilization (50%) with 2 instances of type large.3, and 1
server (server-b) could be under partial utilization (37.5%) with 3
instances of type large.2.
When one of the large.3 instances mapped to server-a is deleted from
physical server type 1, Policy 1 will be violated, since the overall
utilization of server-a falls to 25%.
OpenStack Congress, on detecting the policy violation, uses various
constraint based placement techniques to find the new placement(s)
for physical server type 1 to address the policy violation.
Constrained based placement will be explored in a convex
optimization framework [CONVEX-OPT]; some of the algorithms which
would be considered are linear programming [LINEAR-PROGRAM], branch
and bound [BRANCH-AND-BOUND], interior point methods, equality
constrained minimization, non-linear optimization etc.
Various new placement(s) are described below.
1) New placement 1: Move 2 instances of large.2 to server-a. Overall
utilization of server-a - 50%. Overall utilization of server-b -
12.5%.
2) New placement 2: Move 1 instance of large.3 to server-b. Overall
utilization of server-a - 0%. Overall utilization of server-b -
62.5%.
3) New placement 3: Move 3 instances of large.2 to server-a. Overall
utilization of server-a - 62.5%. Overall utilization of server-b -
0%.
New placements 2 and 3 could be considered optimal, since they
achieve maximal bin packing and open up the door for turning off
server-a or server-b and maximizing energy efficiency.
To detect violations of Policy 1, an example of a classification
rule is expressed below in Datalog, the policy language used by
OpenStack Congress.
Krishnan Expires May 2015 [Page 7]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
Database table exported by the Resource Placement and Scheduler for
Policy 1:
The database table exported by the Resource Placement and Scheduler
for Policy 1 is below.
server_utilization (physical_server, overall_util)
- Each database entry has the physical server and the calculated
average overall utilization.
Policy 1 (in Datalog [DATALOG] policy language):
Policy 1 in a Datalog policy language is as follows.
error (physical_server) :-
nova [OPENSTACK-NOVA-COMPUTE]: node (physical_server, "type1"),
resource placement and scheduler: server_utilization
(physical_server, overall_util < 50)
4. Related Work
A related proof of concept in ETSI NFV on placement and scheduling
is [ETSI-NFV-POC-PLACEMENT].
5. Summary
6. Future Work
TBD
7. IANA Considerations
This draft does not have any IANA considerations.
8. Security Considerations
9. Contributors
Hesham ElBakoury
Huawei
Hesham.ElBakoury@huawei.com
Krishnan Expires May 2015 [Page 8]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
10. Acknowledgements
None.
11. References
11.1. Normative References
11.2. Informative References
[ETSI-NFV-WHITE] "ETSI NFV White Paper,"
http://portal.etsi.org/NFV/NFV_White_Paper.pdf
[ETSI-NFV-USE-CASES] "ETSI NFV Use Cases,"
http://www.etsi.org/deliver/etsi_gs/NFV/001_099/001/01.01.01_60/gs_N
FV001v010101p.pdf
[ETSI-NFV-REQ] "ETSI NFV Virtualization Requirements,"
http://www.etsi.org/deliver/etsi_gs/NFV/001_099/004/01.01.01_60/gs_N
FV004v010101p.pdf
[ETSI-NFV-ARCH] "ETSI NFV Architectural Framework,"
http://www.etsi.org/deliver/etsi_gs/NFV/001_099/002/01.01.01_60/gs_N
FV002v010101p.pdf
[OPENSTACK] "OpenStack Open Source Software,"
https://www.openstack.org/
[OPENSTACK-CONGRESS-POLICY-ENGINE] "A policy as a service open
source project in OpenStack,"
https://wiki.openstack.org/wiki/Congress
[OPENSTACK-CELIOMETER-MEASUREMENT] "OpenStack Celiometer,"
http://docs.openstack.org/developer/ceilometer/measurements.html
[OPENSTACK-NOVA-COMPUTE] "OpenStack Nova,"
https://wiki.openstack.org/wiki/Nova
[LINEAR-PROGRAM] Dmitris Alevras and Manfred W. Padberg, "Linear
Optimization and Extensions: Problems and Solutions," Universitext,
Springer-Verlag, 2001.
[BRANCH-AND-BOUND] "Fundamentals of Algorithmics," G. Brassard and
P. Bratley.
[NFV-MANO-SPEC] "NFV Management and Orchestration Framework
Specification,"
Krishnan Expires May 2015 [Page 9]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
http://docbox.etsi.org/ISG/NFV/Open/Latest_Drafts/NFV-MAN001v061-
%20management%20and%20orchestration.pdf
[CONVEX-OPT] "Convex Optimization,"
https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf
[ETSI-NFV-POC-PLACEMENT] "ETSI NFV Proof of Concept on Placement and
Scheduling,"
http://nfvwiki.etsi.org/index.php?title=Constraint_based_Placement_a
nd_Scheduling_for_NFV/Cloud_Systems
[DATALOG] Ceri, S. et al., "What you always wanted to know about
Datalog (and never dared to ask)," Knowledge and Data Engineering,
IEEE Transactions on (Volume:1 , Issue: 1 )
[IRTF-NFV-POLICY-ARCH] Figueira, N. et al., "Policy Architecture and
Framework for NFV and Cloud Services,"
Authors' Addresses
Ram (Ramki) Krishnan
Brocade Communications
ramk@brocade.com
Norival Figueira
Brocade Communications
nfigueir@Brocade.com
Dilip Krishnaswamy
IBM Research
dilikris@in.ibm.com
Diego Lopez
Telefonica I+D
Don Ramon de la Cruz, 82
Madrid, 28006, Spain
+34 913 129 041
diego.r.lopez@telefonica.com
Steven Wright
AT&T
sw3588@att.com
Tim Hinrichs
VMware
thinrichs@vmware.com
Krishnan Expires May 2015 [Page 10]
Internet-Draft NFVIaaS Policy Resource Placement & Scheduling May 2015
Krishnan Expires May 2015 [Page 11]