Uni Innsbruck Informatik - 1 Grid InterNetworking Michael Welzl DPS NSG Team .

Uni Innsbruck Informatik - Uni Innsbruck Informatik - 11

Grid InterNetworkingGrid InterNetworking

Michael Welzl Michael Welzl http://www.welzl.at

DPS NSG Team DPS NSG Team http://dps.uibk.ac.at/nsgInstitute of Computer ScienceInstitute of Computer ScienceUniversity of InnsbruckUniversity of Innsbruck

GridNets 2006GridNets 2006San Jose, CA USASan Jose, CA USA

1/2 October, 20061/2 October, 2006


OutlineOutline

• Problem scope

• Proposed solutions– Example 1: Network Measurement– Example 2: Grid-Network-Simulation– Example 3: QoS for the Grid

• Conclusion


Problem scopeProblem scope

Shrinking the problem space


What is the Grid?What is the Grid?

• Metaphor: power grid– just plug in, don‘t care where (processing) power comes from,

don‘t care how it reaches you

• Common definition:The real and specific problem that underlies the Grid concept is coordinated resource sharing and problem solving in dynamic, multi institutional virtual organizations[Ian Foster, Carl Kesselman and Steven Tuecke, “The Anatomy of the Grid – Enabling Scalable Virtual Organizations”, International Journal on Supercomputer Applications, 2001]

• Common terms:Virtual Team - members of one or several Virtual Organizations who use a Grid

• Most of the time...– the real and specific goal is High Performance Computing– virtual organizations and virtual teams are well defined

(as opposed to the SETI@Home usage scenario)– i.e. not an „open“ system, security is a big issue


Size

ScopeScope

• Grid history: parallel processing at a growing scale– Parallel CPU architectures– Multiprocessor machines– Clusters– (“Massively Distributed“) computers on the Internet

• Traditional goal: processing power– Grid people = parallel people; thus, goal has not changed much

• Broader definition (“resource sharing“)- reasonable - e.g., computers also have harddisks :-)– New research areas / buzzwords: Wireless Grid, DataGrid,

Pervasive Grid, [this space reserved for your favorite research area] Grid

– sometimes perhaps a little too broad, e.g., “P2P Working Group“ is now part of the Global Grid Forum

Reasonable to focus on

this.


Legacy codes

Components

Web Services

Grid Workflowsbased on activities

MPI HPF OMP

HPFOMPMPI

MPI Java Legacy Codes

Descriptor GenerationComponent Interaction Optimization, Adaptation

Service DescriptionDiscovery, SelectionDeployment, Invocation

Dynamic InstantiationService OrchestrationQuality of Service

Grid Workflow ApplicationsGrid Workflow Applications

• Components are built, Web (Grid) Services are defined,Activities are specified

• Activities (which may communicate with each other) should automatically be distributed by a scheduler


UIBK-DPS development: ASKALONUIBK-DPS development: ASKALONA Grid Application Development and Computing EnvironmentA Grid Application Development and Computing Environment

XML


Grid requirementsGrid requirements

• Efficiency + ease of use– Programmer should not worry (too much) about the Grid

• Underlying system has to deal with– Error management– Authentification, Authorization and Accounting (AAA)– Efficient Scheduling / Load Balancing– Resource finding and brokerage– Naming– Resource access and monitoring

• No problem: we do it all - in Middleware

• de facto standard: “Globus Toolkit“– installation of GT3 in our high performance system: 1 1/2 hours or so...– yes, it truly does it all :) 1000s of addons - GridFTP, MDS, NWS, GRAM, ..– this is just the basis - e.g., ASKALON is layered on top of Globus


Grid-network peculiaritiesGrid-network peculiarities• Special behavior

– Predictable traffic pattern - this is totally new to the Internet!– Web: users create traffic– FTP download: starts ... ends– Streaming video: either CBR or depends on content! (head movement, ..)

• Could be exploited by congestion control mechanisms– Distinction: Bulk data transfer (e.g. GridFTP) vs. control messages (e.g.

SOAP)– File transfers are often “pushed“ and not “pulled“– Distributed System which is active for a while

• overlay based network enhancements possible– Multicast– P2P paradigm: “do work for others for the sake of enhancing the whole

system (in your own interest)“ can be applied - e.g. act as a PEP, ...• sophisticated network measurements possible

– can exploit longevity and distributed infrastructure

• Special requirements– file transfer delay predictions

• note: useless without knowing about shared bottlenecks– QoS, but for file transfers only (“advance reservation“)


Enriched with customisedEnriched with customisednetwork mechanismsnetwork mechanisms

Original Internet technologyOriginal Internet technology

Traditional Internet Traditional Internet applicationsapplications(web browser, ftp, ..)(web browser, ftp, ..)

Real-time multimedia Real-time multimedia applications (VoIP, applications (VoIP, video conference, ..)video conference, ..)

Today‘s Grid Today‘s Grid applicationsapplications

Driving a racing carDriving a racing caron a public roadon a public road

Applications with specialApplications with specialnetwork properties andnetwork properties andrequirementsrequirements

Bringing the Grid to its full potential !Bringing the Grid to its full potential !

EC-GINEC-GIN EC-GINEC-GIN enabled enabledGrid applicationsGrid applications

Research gap: Grid-specificResearch gap: Grid-specificnetwork enhancementsnetwork enhancements


What is EC-GIN?What is EC-GIN?

• European project: Europe-China Grid InterNetworking– STREP in IST FP6 Call 6– 2.2 MEuro, 11 partners (7 Europe + 4 China)– Networkers developing mechanisms for Grids


Research ChallengesResearch Challenges

• Research Challenges:– How to model Grid traffic?

• Much is known about web traffic (e.g. self-similarity) - but the Grid is different!

– How to simulate a Grid-network?• Necessary for checking various environment conditions• May require traffic model (above)• Currently, Grid-Sim / Net-Sim are two separate worlds

(different goals, assumptions, tools, people)

– How to specify network requirements?• Explicit or implicit, guaranteed or “elastic“, various possible levels of

granularity

– How to align network and Grid economics?• Combined usage based pricing for various resources including the network

– What P2P methods are suitable for the Grid?• What is the right means for storing short-lived performance data?


Some issues: application Some issues: application interface...interface...

• How to specify properties and requirements– Should be simple and flexible - use QoS specification languages?– Should applications be aware of this?

Trade-off between service granularity and transparency!

Traditional method Our approach

GIN API

GIN API


... and peer awareness... and peer awareness

(a) Traditional PEP

Intermediary helperGrid end system Grid end system

Data flow

Grid end system

Intermediary helper

Grid end systemData flow Data flow

(b) NSG PEP


Problem: How Grid folks see the Problem: How Grid folks see the InternetInternet• Abstraction - simply use what is available

– still: performance = main goal

Wrong.

• Quote from a paper review:“In fact, any solution that requires changing the TCP/IP protocol stack is

practically unapplicable to real-world scenarios, (..).“

• How to change this view– Create awareness - e.g. GGF GHPN-RG published documents such as

“net issues with grids“, “overview of transport protocols“– Develop solutions and publish them! (EC-GIN, GridNets)

Just like Web Service

community

Absolutely not like Web Service community !

Conflict!

• Existing transport system(TCP/IP + Routing + ..) works well

• QoS makes things better, the Grid needs it!– we now have a chance for that, thanks to IPv6


A time-to-market issueA time-to-market issue

Research begins

(Real-life)coding begins

Real-life tests

begin

Thesis writing

Result: thesis + running code;tests in collaboration withdifferent research areas

Result: thesis + simulationcode; perhaps early real-lifeprototype (if students did well)

Research begins

(Simulation)coding begins

Thesis writing

Typical Grid project

Typical Network project

Ideal


Machine-only communicationMachine-only communication

• Trend in networks: from support of Human-Human Communication– email, chat

• via Human-Machine Communication– web surfing, file downloads (P2P systems), streaming media

• to Machine-machine Communication– Growing number of commercial web service based applications– New “hype“ technologies: Sensor nets, Autonomic Computing vision

• Semantic Web (Services): first big step for supporting machine-only communication at a high level

• So far, no steps at a lower level– This would be like RTP, RTCP, SIP, DCCP, ... for multimedia apps:

not absolutely necessary, but advantageous


The long-term value of Grid-net The long-term value of Grid-net researchresearch

• A subset of Grid-net developments willbe useful for other machine-onlycommunication systems!

Grid

Sensor nets

Web serviceapplications

Future work

• Key for achieving this: change viewpoint from“what can we do for the Grid“ to “what can the Grid do for us“(or from “what does the Grid need“ to “what does the Grid mean to us“)


Proposed solutionsProposed solutions


Example 1: Network MeasurementExample 1: Network Measurement


NWS: The Network Weather ServiceNWS: The Network Weather Service

• Distributed system consisting of– Name Server (boring)– Sensor - actual measurement instance, regularly stores values

in......– Persistent State– Forecaster (calculations based on data in Persistent State)

• Interesting parts:

– SensorMeasured resources: availableCpu, bandwidthTcp, connectTimeTcp, currentCpu, freeDisk, freeMemory, latencyTcp

– ForecasterApply different models for prediction, compare with actual measurement data, choose best match

Duration of a long TCP transfer

RTT of a small message


NWS critiqueNWS critique

• Architecture (splitting into sensors, forecaster etc.) seems reasonable;open source consider integrating new work in NWS

• Sensor– active measurements even though non-intrusiveness was an important design

goal - does not passively monitor TCP (i.e. ignores available data)– strange methodology:

(Large message throughput) “Empirically, we have observed that a message size of 64K bytes (..) yields meaningful results“

– ignores packet size ( = measurement granularity ) and path characteristics– trivial method - much more sophisticated methods

available (e.g. packet pair - later!)– point-to-point measurements: distributed infrastructure not taken into

account

• Forecaster– relies on these weird measurements, where we don‘t know much about the

distribution (but we do know some things about net traffic IFF properly measured)

– uses quite trivial models (but they may in fact suffice...)


Exploiting the Distributed Exploiting the Distributed InfrastructureInfrastructure

• Example problem:– C allocates tasks to A and B (CPU, memory available); both send results to C– B hinders A - task of B should have been kept at C!

• Path changes are rare - thus, possible to detect potential problem in advance– generate test messages from A, B to C - identify signature from B in A‘s traffic

• Another issue in this scenario: how valid is a prediction that A obtains if a measurement / prediction system does not know about the shared bottleneck?

A

B

C


Exploiting longevityExploiting longevity• Time scale of traffic fluctuations < time scale of path changes

knowledge of link capacities may be more useful than traffic estimate

• Underlying technique: packet pair– send two packets p1 and p2 in a row; high probability that p2 is

enqueued exactly behind p1 at bottleneck– at receiver: calculate bottleneck bandwidth via time between p1 and

p2– minimize error via multiple probes– TCP with “Delayed ACK“ receiver automatically sends packet pairs

passive TCP receiver monitoring is quite good!


Traffic prediction by monitoring Traffic prediction by monitoring TCPTCP• TCP propagates bottleneck self-similarity to end systems (“samples bandwidth“)• Automatic prediction? Complex, but possible, I think - e.g.:

Yantai Shu, Zhigang Jin, Jidong Wang, Oliver W. W. Yang: Prediction-Based Admission Control Using FARIMA Models. ICC (3) 2000: 1325-1329

Results from measuring TCP throughput at equidistant intervals

Results from proper TCP monitoring (loss as a congestion indicator)

Available bandwidth

TCP sending rate

Recent related paper (more

realistic, simpler approach):

SIGCOMM 2005


Grid-Network SimulationGrid-Network Simulation


ProcedureProcedure

• Grid simulator only simulates one execution on one machine

• File transfers: generate scenario + invoke network simulator

• Possibility: data transfers influencing each other


Case 1: All tasks and data transfers have equal duration

Grid Simulator / Netzwerk Simulator

Example scenarioExample scenario

Tasks = {T1, T2, T3, T4}Resources = {R1, R2, R3, R4}Data transfers = {D1, D2, D3, D4}


Example scenario /2Example scenario /2

Data transfers with different duration

Grid Simulator / Netzwerk Simulator


ConclusionConclusion

• Implementation in the works

• Method tackles an important and current problem, but...

• Open question: how much time needed between two clusters?– Depends on background traffic and network topology

• Time consuming– Repeated simulation of data transfers– Total # parameters = (# Grid parameters) * (# network

parameters)

• On the other hand...– User = Research group which also runs a Grid– Easy to distribute (parameter study) Grid simulation on the Grid– Seems strange (recursion), but makes sense: one Grid application

for carrying out an analysis with numerous environment conditions


QoS for the GridQoS for the Grid


QoS: the state-of-the-art :-(QoS: the state-of-the-art :-(

Papers from SIGCOMM‘03 RIPQOS Workshop: “Why do we care, what have we learned?“

• QoS`s Downfall: At the bottom, or not at all! Jon Crowcroft, Steven Hand, Richard Mortier,Timothy Roscoe, Andrew Warfield

• Failure to Thrive: QoS and the Culture of Operational Networking Gregory Bell • Beyond Technology: The Missing Pieces for QoS Success Carlos Macian, Lars

Burgstahler, Wolfgang Payer, Sascha Junghans, Christian Hauser, Juergen Jaehnert

• Deployment Experience with Differentiated Services Bruce Davie • Quality of Service and Denial of Service Stanislav Shalunov, Benjamin Teitelbaum • Networked games --- a QoS-sensitive application for QoS-insensitive users?

Tristan Henderson, Saleem Bhatti • What QoS Research Hasn`t Understood About Risk Ben Teitelbaum, Stanislav

Shalunov • Internet Service Differentiation using Transport Options:the case for policy-aware

congestion control Panos Gevros


Key reasons for QoS failureKey reasons for QoS failure

• Required participation of end users and all intermediate ISPs– “normal“ Internet users want Internet-wide QoS, or no QoS at all– In a Grid, a “virtual team“ wants QoS between its nodes– Members of the team share the same ISPs - flow of $$$ is possible

• Technical inability to provision individual (per-flow) QoS– “normal“ Internet users

• unlimited number of flows come and go at any time• heterogeneous traffic mix

– Grid users• number of members in a “virtual team“ may be limited• clear distinction between bulk data transfer and SOAP

messages• appearance of flows mostly controlled by machines, not humans

QoS can work for the Grid !


Proposed architectureProposed architecture

• Goal: efficient per-flow QoS without signaling to routers

• Idea: use traditional coarse-grain QoS (DiffServ) to differentiate between– long-lived bulk data transfer with advance reservation (EF) and– everything else (= SOAP etc. over TCP) (best effort)

• Allows us to assume isolated traffic; planned to drop this requirement later

• Because data transfers are long lived, apply admission control– Flows signal to resource broker (RB) when joining or leaving the network

• Mandate usage of one particular congestion control mechanism for all flows in the EF aggregate– Enables efficient resource usage because flows are elastic


Key ingredients of our QoS soupKey ingredients of our QoS soup

• Link capacities must be known, paths should be stable(capacity information should be updated upon routing change)

• Shared bottlenecks must be known

• Bottlenecks must be fairly shared by congestion control mechanism irrespective of RTT (max-main fairness required, i.e. all flows must increase their rates until they reach their limit)

• No signaling to routers = no way to enforce proper behavior there must be no cheaters– User incentive: fair behavior among cooperating nodes among which

Grid application is distributed– Unfair behavior between Grid application 1 and 2 in same Grid

neglected(usually acceptable, as used by same Virtual Organization)


Link capacities must be knownLink capacities must be known

• Can be attained with measurements• Working on permanently active, (mostly) passive measurement

system for the Grid that detects capacity with packet pair– send two packets p1 and p2 in a row; high probability that p2 is

enqueued exactly behind p1 at bottleneck– at receiver: calculate bottleneck bandwidth via time between p1

and p2– e.g. TCP: “Delayed ACK“

receiver automatically sendspacket pairs passive TCP receivermonitoring is quite good!

– exploit longevity - minimizeerror by listening for along time


Shared bottlenecks must be knownShared bottlenecks must be known

• Simple basis: distributed traceroute tool– enhancement: traceroute terminates early upon detection of known hop

• Handle “black holes“ in traceroute– generate test messages from A, B to C - identify signature from B in A‘s traffic– method has worked in the past: “controlled flooding“ for DDoS detection

A

B

C


Congestion Control mechanism Congestion Control mechanism must be max-min fairmust be max-min fair• Was once said to be impossible without per-flow state in routers

– not true; XCP and some others– but these explicit require router support...

• Main problem: dependence on RTT– three good indications that this can be removed without router support

1. CADPC/PTP (my Ph.D. thesis)...• max-min fairness based on router feedback, but only capacity and

available bandwidth (could also be obtain by measuring)

2. Personal comment by Sally Floyd• Reference to old paper on “phase effects“

3. TCP Libra

• Problem: efficiency - no max-min fair “high-speed“ CC mechanism without router support– searched for a long time– now: plan to change existing one based on knowledge from above

examples

increase/decreasefactors are f(RTT)


Per-flow QoS without signaling to Per-flow QoS without signaling to routersrouters

continuous measurements;update to BB upon path change

Synchronization ofdistributed (P2P based)database; link capacities

known to all brokers1. may I join?2. yes

Synchronization ofdistributed (P2P based)database; all flows known

to all brokers3. I quit4. ok

Traditional method: signaling

to edge routers (e.g. with COPS) at

this point!

Synchronization ofdistributed (P2P based)database; all flows known

to all brokers


Efficiency via elasticityEfficiency via elasticity

• QoS guarantees in Grid: „File will be transferred within X seconds“ enables flexible resource usage

Time (seconds)

Flow 1

Bottleneck Bandwidth (kbs)

4 kbs

End 1 End 2 End 3 End 4

Flow 2

Flow 3

Flow 4


Efficiency via elasticity /2Efficiency via elasticity /2

• Flow 1 stopped, flows 2-4 automatically increase their rates– leading to earlier termination times E2‘-E4‘; known to (calculated by) BB

Time (seconds)

Flow 4


4 kbs

t1 - E1 E2 E3 E4 E2’ E3’ E4’

Flow 3

Flow 2



• Flow 5 asks BB for admission– BB knows about current rates and promised E2-E4, grants access

Time (seconds)


Flow 4

4 kbs

t2 – Flow 5 wants to get admission here

E2 E3 E4 E2’ E3’ E4’

Flow 3

Flow 2



• Flow 2 terminates in time– Flows 3-5 will also terminate in time

Time (seconds)

Flow 4


4 kbs

t2 – Flow 5 gets admission here

E2 E3 E4 E2’ E3’ E4’

Flow 3

Flow 2

E5

Flow 5

E4” E3” E2”

Additional flow admitted and earlier termination times than promised!


Elasticity without Congestion Elasticity without Congestion Control?Control?

• Significant amount of additional signaling necessary

Time

Flow-1

Flow-2

Flow-3

Flow-4

Bottleneck Bandwidth

4 kbs

As Flow-1 stops, Flows 2-4 could increase their

rates

Without congestion control, signal “increase your rates“ to flows 2-4

required!

As flow 5 is admitted, signal “reduce your

rates“ toflows 2-4 required!


Additional considerationsAdditional considerations

• How to assign different rates to different flows?– max-min fairness: if a sender “acts“ like two, it obtains twice the

rate– consider rate consisting of slots (e.g. 1 kbit/s = 1 slot)– flows can consist of several slots– let congestion control mechanism operate on slots

• Possibility: admit new flows even in scenario below

Time (seconds)

Flow 4


4 kbs

t2 – Flow 5 gets admission here

E2 E3 E4 E2’ E3’ E4’

Flow 3

Flow 2

E5

Flow 5

E4” E3” E2”

Must introduce unfairness: only flow 2

can reduce rate

Disadvantage: more signaling again!


Difficult & distant future workDifficult & distant future work

• Drop requirement of traffic isolation via DiffServ– constantly obtain and update conservative estimate of available

bandwidth using packet pair (works without saturating link)– ensure that limit is never exceeded; “condition red“ otherwise!– Some open questions...

• does this require the CC mechanism to be TCP-friendly?• condition red: reduce slots, or let flows be aggressive for a short

time?

• How to handle routing changes– will be noticed, but can reduce capacity break QoS guarantee– condition red; can happen in worst case, but to be avoided at all cost– mitigation methods

• very conservative estimate of available bandwidth; leave headroom• tell senders to reroute via intermediate end systems

• Bottom line: lots of complicated issues, but possible to solve them



• Grid applications show special requirements and properties from a network perspective– and it is reasonable to develop tailored Internet technology for them.

• There is another class of such applications...

• Multimedia.

• For multimedia applications, an immense number of network enhancements (even IETF standards) exist.

• For the Grid, there is nothing.

• This is a research gap; let‘s fill it together!– submit a paper to GridNets 2007 !

Reminder: if done right, such research is also applicable to other

systems with machine-only traffic


Thank you!Thank you!

Questions?

Uni Innsbruck Informatik - 1 Grid InterNetworking Michael Welzl DPS NSG Team .

Documents

Transcript of Uni Innsbruck Informatik - 1 Grid InterNetworking Michael Welzl DPS NSG Team .