november 201079 th ietf @ beijing, china1 cross stratum optimization (cso) bar bof
TRANSCRIPT
November 2010 79th IETF @ Beijing, China 1
Cross Stratum Optimization (CSO) Bar BOF
November 2010 79th IETF @ Beijing, China 2
Agenda
• IETF CSO Effort Background (Young Lee, 5 min)• NA-ARAM (Ning So, 15 min)• NS Query (Young Lee, 15 min)• Q & A (25 min)
November 2010 79th IETF @ Beijing, China 3
IETF CSO Effort• CLO Bar BOF was held in Maastricht IETF meeting (78th)
– 40 + attendants from carriers, vendors, academia and research institutes
• Created IETF mailing list: [email protected] • Public Archive: http://www.ietf.org/mail-archive/web/cso• CLO -> CSO (Name changes) • CSO := {NS Query, NA-ARAM}
Where NS Query = Network Stratum QueryNA-ARAM = Network Aware Application Resource Assignment and Mobility
• NS Query BOF Request didn’t go through the initial IESG approval– Need more clarification
November 2010 79th IETF @ Beijing, China 4
Network Aware Application Resource Assignment and Mobility in Data Centers
draft-so-network-aware-application-problem-01.txt
Ning So (UTD) Young Lee (Huawei) Dave McDysan (Verizon) Greg Bernstein (Grotto)Tae Yeon Kim (ETRI) Kohei Shiomoto (NTT)Oscar Gonzalez de Dios (Telefonica)
November 2010 79th IETF @ Beijing, China 5
Problem Description• Many application services offered by Data Center
make significant use of the underlying networks resources
• The same application service can be offered by multiple geographically dispersed data centers. – allowing the application to be closer to the end-users to
enhance service performance and user experience • Problem #1: The decision as which server/data
center to select for an application request from end-users can negatively affect the quality of experience (QoE) of the users if not done correctly.
November 2010 79th IETF @ Beijing, China 6
Problem Description (Cont’d)• VMs supporting the applications can be actively
distributed within and/or between data centers. – Improve server utilization– Prevent service performance degradation during the peak
usage time period, and during equipment failures. • Problem #2: When instantiating new VMs or
migrating existing VMs across data centers, the underlying network loading conditions within a data center (LAN) or between data centers (MAN/WAN) need to be considered
November 2010 79th IETF @ Beijing, China 7
Exemplary Cloud Data Center
End-User 1
1 2
3
CE 1
CE 2
12
PE 1
PE 2
PE 4
PE 3
PE
1 2
CE 3
ApplicationServers
CE 4
AccessNetwork Carrier
BackboneNetwork
DC 1
DC 2
DC 3AccessNetwork
PE
CE 4End-User 2
Global Load Balancer
November 2010 79th IETF @ Beijing, China 8
Current Server Selection Technology
• The server selection for an application/VM is done by load-balancer. – The load balancer is aware of a certain level of
server usage data and distributes the application requests based on that data.
November 2010 79th IETF @ Beijing, China 9
Deficiencies of Current Server Selection Technology
• The current load balancing technology is insufficient in providing an optimal decision across multiple VLANs and multiple Data Centers – No standard solution for the communication exchange among load
balancers located in different Data Centers • load balancers from different vendors cannot communicate to each other
– load balancers know little about the underlying network conditions• When migrating existing VMs/applications from one data center to
another, the underlying network load condition in LAN/MAN/WAN can be constraining factors.
– Load balancers know little about user condition
November 2010 79th IETF @ Beijing, China 10
Optimized Server Selection Criteria
• Factors need be considered in choosing the right server in the right data center for an application request or in instantiating new or migrating existing VMs/applications– Server level load condition in a data center– Intra data center network condition – Carrier MAN/WAN network condition – User Condition
November 2010 79th IETF @ Beijing, China 11
Server Selection Criteria Details• Server level load condition
– VM supportability limitations – CPU utilization– Memory segmentation and consumption– Application limitations e.g. max # of simultaneous instances of the application
supportable– Storage access speed (disk, RAM, etc.)– Environment considerations such as server temperature, power load, and
electrical cost at the time– Operational and managerial considerations such as location of peer servers
and storages– VM to NIC switching in a virtual switch
• Intra-DC network condition – Server NIC to Top of Rack (ToR) Switch– TOR switch to Layer 2 Switch - link and node level– Between L2 Switches and L2 switch to L2 core/gateway switch/router - link
and node level– L2 gateway router to provider edge (PE) router - link and node level.
November 2010 79th IETF @ Beijing, China 12
Server Selection Criteria Details• Carrier MAN/WAN network condition
– Type of networks and the technical capabilities of the networks– Bandwidth capabilities and availability – Latency– Jitter– packet loss– other Network Performance Objective (NPO) as defined in section 5 of
[ITU-T Y.1541] • User condition
– User access capabilities and limitations (e.g., user terminal information such as codec for video application)
– User location– Optional user preferences (for some application, user may be able to
specify its preferences. For example, the preferred server location for gaming).
November 2010 79th IETF @ Beijing, China 13
Requirements for Optimized NA-ARAM• End-users send requests to the Application Controller (global load
balancer), which makes server assignment decision for the user application.
– End-users may send the following• Required application: this may be a simple URL request • Specific server identity or location. • Additional security requirements• Modified application specs. (for mobile users) • Optional end-user terminal information
• Inter DC communication: the application controller need to be aware – Server/Application Status from each of the Data Centers where the
application servers are located – Intra-DC network conditions from each of the Data Centers where the
application servers are located• Data Center-Network Stratum Communication
– The details of how this can be accomplished are addressed in Network Stratum Query draft.
November 2010 79th IETF @ Beijing, China 14
Thanksand
Questions?
November 2010 79th IETF @ Beijing, China 15
NQ Query
draft-lee-network-stratum-query-problem-01.txt
Young Lee (Huawei) Dave McDysan (Verizon)Ning So (UTD) Greg Bernstein (Grotto)Tae Yeon Kim (ETRI) Kohei Shiomoto (NTT)Oscar Gonzalez de Dios (Telefonica)
November 2010 79th IETF @ Beijing, China 16
Backgrounds• Most Internet Services are offered by the “servers”
geographically spread. • Data Centers Networks have emerged as physical
infrastructure for Content Delivery Networks and Cloud Computing.
• Application services, e.g., data centers, make significant use of the underlying network services
• Application services have access to little or no information about network services, e.g., available bandwidth, congestion level, etc.
November 2010 79th IETF @ Beijing, China 17
Context of NS Query:Data Center Networks (DCN)
End-User 1
GLB 1 2
3
GR
GR
12
PE 1
PE 2
PE 4
PE 3
PEGLB
1 2
GR
GLB
ApplicationResource(server)
CE 4
AccessNetwork Carrier
BackboneNetwork
DC 1
DC 2
DC 3
(3) Inter DC Communication(exchange server performance data)
(1) End-user to DC communication(request/reply)
(4) DC-Carrier Communication(NS Query)
AccessNetwork
PE
CE 4
End-User 2
(2) Intra DC Communication
Direct Access(Corporate User)
CE 4
November 2010 79th IETF @ Beijing, China 18
Cross-Stratum Application Stratum• Distributed Resources: Data Centers
with servers, content, data sets, computing power, cache/mirror
• Uses Network Resources• Different QoS requirements for each
application
Network Stratum• Bandwidth, Connections, Links, • Connection Processing (Creation,
Deletion, Management)• Admission Control, Resource Reservation • Applications uses resources in IP, MPLS,
and/or Optical Transport Networks, Layer 2
Network Stratum
Application Stratum
NS Query
November 2010 79th IETF @ Beijing, China 19
Application Service ProfilesCharacteristics & QoS Requirement of application service from a
network perspective:• Location profile: locations of both the clients and the sources• QoS profile: (i) Delay Tolerance Bound; (ii) Jitter Tolerance Bound; (iii) Packet Delivery Ratio
Tolerance; (iv) Network Availability, etc.• Connectivity profile: (i) P-P; (ii) P-MP; (iii) MP-MP; (iv) Any Cast• Directionality profile: (i) uni-directional; (ii) bi-directional• Bandwidth profile: Maximum, average, and minimum bandwidth requirements for the connectivity,
maximum burst rate, maximum burst duration, etc.• Duration of service profile: service time of the application• Network media profile: (i) optical only; (ii) no microwave, etc.• Restoration profile: (i) Reroute required; (ii) do not re-route, etc. • Security profile: (i) dedicated end-to-end VPN-like resource allocation; (ii) dedicated physical
resource allocation
Currently, network is not informed of the nature of the application --- there is no good way to convey the application service profile to network stratum
November 2010 79th IETF @ Beijing, China 20
Carrier MAN/WAN network condition
• Type of networks and the technical capabilities of the networks;
• Bandwidth capabilities and availability;• latency;• jitter; • packet loss;• And other Network Performance Objective
(NPO) as defined in section 5 of [ITU-T Y.1541].
November 2010 79th IETF @ Beijing, China 21
CSO NS Query ArchitectureGlobalLoad
Balancer
ApplicationControl
Gateway
Inter DCGateway
Intra DCResource
RemoteData Centers
Interfaces
End-UserInterfaces
NetworkControl
Gateway
NS Query (First Step)
IPPM
TED
LSDBBGP-RIBIGP-RIB
NS Query (Second Step)
November 2010 79th IETF @ Beijing, China 22
Two-stage Query• A Vertical Query
– First Stage: A vertical query from an application entity (i.e., the Application Control Gateway (ACG) in Data Center) to an entity representing the network (i.e., the Network Control Gateway (NCG)) for highly summarized and abstracted network related information for a particular point in network; and
• A Horizontal Query– Second Stage: Internal "horizontal query" at the network layer along with
summarization and abstraction of the network information in a form that preserves network confidentiality and significantly reduces the amount of information that needs to be transferred.
– The raw information needed to perform this summarization/abstraction is defined in existing and emerging network management standards and protocols (SNMP, Netflow, sFlow, IPPM, IGP, RIB, etc...).
– NS Query would not necessarily standardize how this "internal horizontal queries" and summarization would be performed but would illustrate how such processes can be accomplished via standards, emerging standards or common commercial practices.
November 2010 79th IETF @ Beijing, China 23
NS Query Example• For a particular point in networks, the NS Query can ask the
following network load data in a summarized/abstract form:– b/w availability (this case the minimum b/w requirement should be
provided) – latency – jitter – packet loss
• Note that this can be asked in a different way. For example, the query can simply ask: – Can you give me if you can route x amount of b/w (from server to
end-user) within y ms of latency? – Can you give me if you can route x amount of b/w (from server to
end-user) with no packet loss?
November 2010 79th IETF @ Beijing, China 24
Thanks!