P2P technologies, PlanetLab, and their relevance to Grid work. Matei Ripeanu The University of Chicago

Transcription

1 P2P technologies, PlanetLab, and their relevance to Grid work Matei Ripeanu The University of Chicago

2 Why don t we build a huge supercomputer? Top500 supercomputer list over time: Zipf distribution: Perf(rank) rank -k LinPack perf.gflops (log scale) Parameter 'k' evolution Rank (log scale)

3 Impact Trend: it is increasingly interesting to aggregate the capabilities of the machines in the tail of this distribution. A virtual machine that aggregates the last 10 in Top500 would rank 32 nd in 95 but 14 th in 03 Both Grid and P2P computing are results of this trend: Grids: focus on assembling (a relatively small number of) resources to enable controlled, secure resource sharing P2P focus: scale, deployability. Parameter 'k' evolution. Challenge: design services that offer the best of both worlds complex, secure services, that deliver controlled QoS; are scalable and can be easily deployed

4 Roadmap P2P Definitions Impact Applications Mechanisms PlanetLab Why is this interesting for you? Compare resource management in PlanetLab and Globus Wrap-Up

5 P2P Definition(s) Def 1: A class of applications that take advantage of resources (e.g., storage, cycles, content) available at the edge of the Internet. Edges often turned off, without permanent IP addresses, etc. Def 2: A class of decentralized, self-organizing distributed systems, in which all or most communication is symmetric. Lots of other definitions that fit in between

6 P2P Impact (1) Widespread adoption leading to KaZaA 170 millions downloads (3.5M/week) one of the most popular applications ever! (almost) zero-cost data distribution is forcing companies to change their business models FastTrack might impact copyright laws imesh edonkey DirectConnect Blubster Gnutella Cvernet 4,114,120 1,375, , ,552 97,128 92,678 91,750 Source: 2/4/2003

7 P2P Impact (2) Killer application for broadband to consumers P2P generated traffic may be the single largest contributor to Internet traffic today 100% 80% 60% Internet2 traffic statistics Other Data transfers Unidentified File sharing 40% 20% 0% Feb.'02 Aug.'02 Feb.'03 Aug.'03 Feb. '04 July'04 Source:

8 P2P Impact (3) A huge pool of underutilized resources lays around, users are willing to donate these resources Seti@Home $1.5M / year in additional power consumed which can be put to work efficiently (at least for some types of applications) Users Results received Total CPU Time Floating point operations Total 4,236, M 1.3M years Last 24 hours 2, M 1.3K years 51.4 TFLOPS Source: Seti@Home website, Oct. 2003

9 Roadmap P2P Definitions Impact Applications Mechanisms PlanetLab Why is this interesting for you? Compare resource management in PlanetLab and Globus Wrap-Up

10 Applications: Number crunching Examples: Entropia, UnitedDevices, DistributedScience, many others Approach suitable for a particular class of problems. Massive parallelism Low bandwidth/computation ratio Error tolerance Problems: Centralized. How to extend the model to problems that are not massively parallel? Relevant to Grid space: Valuable resources at the network edge Shows ability to operate in an environment with limited trust in participating resources

11 Applications: File sharing The killer application to date Too many to list them all: Napster, FastTrack (KaZaA, imesh), Gnutella (LimeWire, Morpheus, BearShare), Relevant to Grid space: Building a reliable, high-performance (?) data-delivery service using a large set of unreliable components.

12 Applications: Content Streaming Streaming: the user plays the data as as it arrives Possible solution: The first few users get the stream from the server New users get the stream from the server or from users who are already receiving the stream P2P approach source Oh, I am exhausted! Client/server approach Relevant to Grid space: offload part of the server load to consumers to improve scalability

13 Applications: Performance benchmarking Problem: Evaluate the performance of your Web site form end-user perspective Multiple views on your site performance Generate Internet statistics Connectivity statistics Routing errors, routing maps Relevance to Grid space: Grid clients are heterogeneous, geographically dispersed How do I benchmark my services for this set of consumers?

14 End-to-end Performance Benchmarking Back-end Infrastructure Network Landscape Last-mile Blind Spot Database App server Web server 3 rd party content Firewall Backbone Backbone Regional Network ISP ISP Major Provider Enterprise Provider Local ISP T1 Corporate Network Corporate User Component Testing Datacenter Monitoring Consumer User End-to-end Web Performance Testing Slide source:

15 Many other P2P applications Backup storage (HiveNet, OceanStore) Collaborative environments (Groove Networks) Instant messaging (Yahoo, AOL) Web serving communities (userv) Spam filtering Anonymous Censorship-resistant publishing systems (Ethernity, Freenet)

16 Mechanisms To obtain a resilient system: integrate multiple components with uncorrelated failure curves, and use data and service replication. To improve delivered QoS: move service delivery closer to the user, and integrate multiple providers with uncorrelated demand curves (reduces over-provisioning for peak loads) To generate meaningful statistics, to detect anomalies: Provide views from multiple vantage points To provide anonymity: use large number of nodes, hide in the crowd, and make search costly (or impossible)

17 Roadmap P2P Definitions Impact Applications Mechanisms PlanetLab Why is this interesting for you? Compare resource management in PlanetLab and Globus Wrap-Up

18 PlanetLab Testbed to experiment with your P2P/Grid applications. >400 nodes, >150 sites, 452 nodes 162 sites 450 research projects Funding: Intel, HP, Princeton University, View presented to users: a distributed set of VMs Allocation unit: a slice = a set of virtual machines (VM), one VM at each node. Slice K Slice K Slice K Slice K OS OS OS OS

19 PlanetLab usage examples Stress-test Globus Replica Location Service GSLab: a playground for experimenting with grid-services users have N machines already configured with Globus one mouse-click away Better-than-Internet services: multipath TCP (mtcp) Resilient Overlays Multicast Overlays

20 Why should you find PlanetLab interesting? 1. Open, large-scale testbed to test your P2P applications or your Grid services 2. Solves a similar problem to Grids/Globus: building virtual organizations (or resource federations) Grids: testbeds (deployments of hardware and software) to solve computational problems. PlanetLab: testbed to play with CS applications Main problem for both: enable resource sharing across multiple administrative domains

21 But Roadmap: Attempt to trace differences back to starting assumptions on: user communities, applications, resources. Look at mechanisms to build VO.

22 Assumptions: User communities PlanetLab: users are CS scientists that experiment with and deploy infrastructure services. Globus: users come from a more diverse pool of scientists that are interested to run efficiently their (enduser) applications. Implication: functionality offered User applications Globus OS User applications PL Services PlanetLab. OS

23 Assumptions: Application characteristics Different view on geographical resource distribution: PlanetLab services: «distribution is a goal» leverage multiple vantage points for network measurements, exploit uncorrelated failure rates in large sets of components, Grid applications: «distribution is a fact of life» generally resource distribution is a fact of life: a result of how the VO was assembled (due to administrative constraints). Implication: mechanism design for resource allocation

24 Assumptions: Resources PlanetLab mission as testbed for a new class of networked services allows for little HW/SW heterogeneity. Globus is supported for large set architectures; for sites with multiple security requirements Implications: complexity, development speed

25 Assumptions: Resource ownership PlanetLab Goal: individual sites retain control over their resources PlanetLab limits the autonomy of individual sites in a number of ways: VO admins: Root access, Remote power button Sites: Limited choice of OS, security infrastructure Globus imposes fewer limits on site autonomy Requires fewer privileges (also can run in user space, ) PlanetLab emphasizes global coordination over local autonomy to a greater degree than Globus Implications: ease to manage and evolve the testbed Control at VO level Globus Individual site autonomy

26 Building Virtual Organizations Individual node/site functionality Mechanisms at the aggregate level Security infrastructure Delegation mechanisms Resource allocation and scheduling Resource discovery, monitoring, and selection.

27 Delegation mechanisms: Identity Delegation Broker/scheduler usage scenario: User A submits a job to S1 which, in turn, submits to S2. S2 needs to do accounting or make authorization decisions based on the identity that originated the job (A). Globus Implementation based on delegated X.509 proxies PlanetLab None

28 Delegation mechanisms: Delegating the Right to Resources Broker/scheduler usage scenario: User A acquires capabilities from various brokers then submits the job. Globus/GGF WS-Agreements protocols: To represent contracts between providers and consumers. Local enforcement mechanism is not specified The two efforts are complementary! PlanetLab Individual nodes managers hand out capabilities: akin to timelimited reservations Capabilities can be traded Extra layer to: provide secure transfer, prevent double spending, offer external representation

29 Global resource allocation and scheduling Globus PlanetLab Users Application Managers Brokers / Agents Node Managers Nodes Identity delegation Sends job descriptions Resource usage delegation Sends capabilities (leases)

30 Functionality & infrastructure Wrap-up Grids Convergent Environment: Large, Dynamic, Self-Configuring Grids Large scale: o Weak trust assumptions o Ease of integration Intermittent resource participation Little technical support Diversity in shared resources Build infrastructures to support diverse applications P2P Scale & volatility

31 What can Grids learn from P2P? Scalability Light-weight implementations Aggregate desktops and other small resources Intermittent operation, highly dynamic connectivity Autonomy

32 More information Links to papers, tech-reports, slides: Thank you.