1 working group on ad-hoc networks reliable server pooling for ad-hoc networks m. Ümit uyar, city...
TRANSCRIPT
1
Working Group on Ad-Hoc Networks
Reliable Server Pooling for
Ad-Hoc Networks
M. Ümit Uyar, City College of CUNY
Mariusz Fecko, Telcordia Technologies, Inc.
Murray Hill, NJ
February 5, 2003
Prepared through collaborative participation in the Communications and Networks Consortium sponsored by the U.S. Army Research Laboratory under the Collaborative Technology Alliance Program, Cooperative Agreement DAAD19-01-2-0011
2
Reliable server pooling (RSP)
Objective: Increase system availability by viewing a pool of redundant information sources as a single transport endpoint
Benefits:• clients can get service even if one or more servers fails• underutilized servers can provide relief to nearby overloaded servers• switchover when QoS drops; sustaining sessions through switchovers
Client view:pool user (PU)
(client) server pool ID: pool handle
pool user(client)
Actual systempool elements (PEs)
PE
PEPE PEPE
PE
client’s RSP layer requests
pool handle/pool elements
mapping from its Home Name Server
3
RSP architecture and protocol stack
Key components and issues• signaling protocols• server selection• transport and security state transfer (?)• load sharing and server fault management• transparent connection/session migration (?)
PE(i) PE(j)
PE(k) PE(l)
… …
NS
PU
(4)(2)
(3)
(1)Applications:
RSP-blindRSP-partially-aware
RSP-fully-aware
RSP API
RSP
Mapping (shim)
SCTP/TCP/UDP
4
RSP scope
Loosely-coupled, open architecture• high survivability: server pools may span several network domains• does not attempt to provide complete fault-tolerant computing
Defining properties• maps a single communication-destination name to multiple routable and
reachable transport endpoints to increase the availability of distributed software-server entities registered under that name
• reliability services, ranging from simple server selection to a fully automatic session-failover capability, can be designed and added using RSerPool as a distributed, open platform
Outside core scope, but work is ongoing• fully transparent session-failover capability• keeping data integrity among servers• data replication techniques in ad-hoc networks
5
RSP in ad-hoc networks
IETF RSerPool WG is defining an architecture and protocols for RSP• Endpoint Name Resolution Protocol (ENRP) and Aggregate Server
Access Protocol (ASAP)• management and maintenance of the entire (flat) server pool
namespace provided by Name Servers (NSs)• lightweight, best-effort design
Is emerging IETF RSerPool applicable to ad-hoc networks?• IETF focuses on fixed, wired networks, and telephony applications• nodes are highly mobile, have limited processing power, are capable of
rapid self-configuration and autoaddressing of nodes and interfaces• is best-effort, lightweight approach sufficient for server resources that
are constrained and dynamic in MANETs?• mobility issues: network partitioning, local vs global name spaces,
dynamic vs static home server• impact on lower layers (routing, MAC) needs to be considered
6
Our approach
Evaluate IETF architecture and protocols through simulation• performance, scalability, resilience to stress conditions
Equip the architecture with mobility solutions • replace fixed elements with dynamic ones (e.g., home name server)• introduce local name spaces to tackle network partitioning• eliminate excessive signaling overhead by making the protocols
mobility-aware
Introduce “enhanced” architecture• distributed server selection and session-admission control• membership of pools based on managing dynamic server resources
•degree of readiness (DR): the number of associations that a server (from a pool) is willing to sustain at a certain level of QoS
• server may belong to different server pools, which can be formed (or taken down or merge/split) dynamically, at varying DRs
7
Simulation capabilities
Realistic evaluation must be performed in multi-layer wireless system
CCNY developed ns-2 simulation testbed that includes:• full implementation of IETF RSerPool protocols (ENRP + ASAP)• integration of MAODV for multicast• movement pattern selection (frequency, speed)• transmission pattern selection (antenna model, transmission range)• ongoing or planned
•group mobility models (Reference Point Group Mobility, Mobility Vector)• integration with UD SCTP ns-2 model
ENRPAgent
ASAPAgent
MAODVRouting
SimplifiedSCTP
RSP-awareCBR APP
Deterministic/StatisticalError Model
Wired/WirelessScenarios Definition
NS-2
8
Results for the IETF RSerPool
Hunting for home name server Successful transmissions
among Name Servers
Ratio of data packets to
control messages
Everything works fine for wired and stable networks; not so for mobile
and dynamic ones
9
Evaluation of IETF approach
Problems for wireless and mobile• namespace synchronization: name servers communicate successfully only 3-18% (90% for
wired) due to network partitioning• huge signaling overhead: only 3-6 as many data packets as control messages (1,300 for
wired)•hunting for home server accounts for 50% of control messages•exchange of info among NSs for 25%
• false server deregistrations: temporary movement of operational server out of transmission range interpreted as failure
Alternative mechanisms designed and evaluated:• replace hunting for home name server with advertisements
•overhead reduction ranges from 22% to 28%• make PE failure detection less aggressive (heartbeats, distinguishing between failed and
temporarily out of range)
10
Alternative mechanism: Server heartbeats
Original mechanism
(3) (1)
PU1
NS1(PU1’s home)
PE2PE1
Pool
(4)
(5)(2)
NS2(PE2’s home)
(4) (2)
PU1
NS1(PU1’s home)
PE2PE1
Pool
(1)
(5)(3)
NS2(PE2’s home)
(6)
(6) (7)
60-100% reduction in false server deregistrations
PE heartbeat mechanism
11
Quantifying RSP benefits
Lost (a) vs (a,b) sustained sessions
Session Sustainability Gain (SSG): relative increase in the system's ability to admit and run a session until completion
SSG measurements for mobile networks:• assumption: at each switchover, we can migrate the session• "binary" sustained vs lost: 10-29%• "reclaimed" session lifetime: 23-42%
(a)
tf
tc(b)
tf
tcf1 f1 f2 f3
tl
(c)
tf
tc
f1 f2 f3 f4
tl=tf
12
Dynamic (re-)configuration
Management of pool resources based on• periodic exchanges of aggregated information and requirements of the
pools• information about the network health from bandwidth brokers and
mobility events from mobility agents• server X advertises its capability set as <ST, SP, DR>, where ST is
server type, SP is server priority, and DR is degree of readiness
Example• server X sends periodic advertisements {<SIP,1,1000><HTTP,2,500>}• DRSIP=1000 and DRHTTP=500 can change between advertisments,
depending on changing server characteristics such as battery life, etc.• DRs are dynamically adjusted and used to perform admission control
based on server resources• server resources (I.e., server membership in pools) readjusted
dynamically to provide end-to-end quality for the service being requested by PU
13
Server selection (distributed, QoS-based)
In resource-constrained environment, the best-effort policy to offer server pool resources may lead to • degradation of service to the already existing sessions• increase in the number of switchovers if individual servers become
overloaded• impaired ability to perform successful switchovers or start new sessions
if server pools become overloaded• best-effort policy (in other words, common sharing policy) results in a
very high number of unsuccessful session attempts when the network resources are stressed [TON2001-Beard]
Four approaches to server selection are being investigated• pool guard channels• utility-based server-resources allocation• prioritized server-resources allocation• distributed fault localization techniques
14
Example Network
R
AP
PEPE
PE
R R
R
APPE
PE
PE
R
hub
PE PU
R
AP
PE
PU
ENRP
ENRP
ENRP
ENRP
MANET 1 MANET 2 MANET 3
PU
PU
PU
PUPU
Backbone Network
ENRP Name Server(Endpoint Name Resolution Protocol)R Router
PU Pool User (Host/Client)hub Hub
AP Access PointPE PE Pool Elements (each in different pools)
single node that is PE in two poolsPE PE
PU
PE PE
15MANET 2MANET 1
R
R R
R
R
hub
PE PU
R ENRP
ENRP
ENRP
ENRP
Backbone Network
APPE
PE
PU
PU
Experiment Scenario 1: PE movement
PEPE
PU
MANET 3
PE
The PE for the “pink” pool moves from MANET 2 to 1• PE is assigned a new IP address• PE must reregister with the ENRP server. • Sessions in progress at PUs rely on SCTP mobility to maintain session• PUs in pink pool may need to switchover to a new PE for new sessions
The PE for the “pink” pool moves from MANET 2 to 1• PE is assigned a new IP address• PE must reregister with the ENRP server. • Sessions in progress at PUs rely on SCTP mobility to maintain session• PUs in pink pool may need to switchover to a new PE for new sessions
PE
APPE
PUPU
PU
SCTP mobility used for in-progress session
RSerPool switchoverused for new sessions
PU
PE PE
AP
16
R
R R
R
R
hub
PE PU
R ENRP
ENRP
ENRP
ENRP
Backbone Network
APPE
PE
PU
Experiment Scenario 2: PU movement
MANET 2
AP
PE
PU
MANET 1
A PU using both services moves from MANET 1 to 2• PU is assigned a new IP address• Transactions in progress at PU rely on SCTP mobility to maintain session • PU may need to switchover to new PEs in green and pink pool,
and may need to select a new primary ENRP server.
A PU using both services moves from MANET 1 to 2• PU is assigned a new IP address• Transactions in progress at PU rely on SCTP mobility to maintain session • PU may need to switchover to new PEs in green and pink pool,
and may need to select a new primary ENRP server.
PE
PU
MANET 3
APPE
PUPU
PUPE
PU
PE PE
17
R
R R
R
R
hub
PE PU
R ENRPENRP
ENRP
ENRP
Backbone Network
APPE
PE
PU
Exp Scenario 3: Failure of co-located Rtr/ENRP
MANET 2
AP
PU
MANET 1
PU
Router and ENRP server co-located on single assetFailure of that asset causes AP to failover to backup
connection to router (e.g. on a lower bandwidth radio channel)PEs in MANET 1 must failover to backup ENRP serverWhat is the effect on PUs, PEs in MANET 1?
Router and ENRP server co-located on single assetFailure of that asset causes AP to failover to backup
connection to router (e.g. on a lower bandwidth radio channel)PEs in MANET 1 must failover to backup ENRP serverWhat is the effect on PUs, PEs in MANET 1?
PE
PE PE
MANET 3
APPE
PUPU
PUPE