1. JuxMem: An Adaptive Supportive Platform for Data Sharing on the Grid Gabriel Antoniu, Luc Bougé, Mathieu Jan IRISA / INRIA & ENS Cachan, France Workshop on Adaptive Grid Middleware New Orleans, September 2003

2. 2 Context: Data Management on the Grid Distributed numerical simulations (code coupling) Problem: data management Solid mechanics Thermodynamics Optics Dynamics Satellite design

3. 3 Existing Data Management Systems Non-transparent large scale data management GridFTP (Globus) and MPI-IO Security, heterogeneity Internet Backplane Protocol (IBP) Control Explicit transfer No consistency guarantee

4. 4 Existing Data Management Systems Transparent small-scale data management Distributed shared memory (DSM) Transparent access Transparent data localization Consistency models and protocols Static, homogeneous architecture

5. 5 Another Approach: Peer-to-Peer Systems Peer-to-peer systems (P2P) Distributed (large-scale) Volatile peers Peers have the same capacities and responsibilities Sharing immutable data Centralized (Napster) Flooding (Gnutella, KaZaA) Distributed hash table (CFS, PAST) Sharing mutable data One writer per data, static assumptions (OceanStore) Manual conflict resolution (Ivy)

6. 6 DSM systems and P2P systems Comparing basic hypotheses DSMP2P Scale10 1 -10 2 10 5 -10 6 DynamicityNullHigh Resource homogeneity Homogeneous (clusters) Heterogeneous (Internet) Control and trust HighLow TopologyFlat Data typeMutableImmutable Typical applications Scientific computation File sharing and storage

7. 7 Idea: Data Sharing Service Proposal: hybrid approach DSM systems: consistency and transparent access P2P systems: scalability and high dynamicity DSMGrid Data ServiceP2P Scale10 1 -10 2 10 3 - 10 4 10 5 -10 6 DynamicityNullMediumHigh Resource homogeneity Homogeneous (clusters) Rather heterogeneous (clusters of clusters) Heterogeneous (Internet) Control and trust HighMediumLow TopologyFlatHierarchicalFlat Data typeMutable Immutable Typical applications Scientific computation Scientific computation and data storage File sharing and storage

8. 8 A Data Sharing Service for the Grid Internet Persistence

9. 9 A Data Sharing Service for the Grid Internet Data transfer ? Transparent data location

10. 10 A Data Sharing Service for the Grid Internet Scalability Internet

11. 11 A Data Sharing Service for the Grid Internet Volatility tolerance

12. 12 JXTA: a Framework for P2P Services Open-source platform for programming P2P applications http://www.jxta.org A peer Uniquely identified (ID) Address independent of physical location Multiple network access points (TCP, HTTP, etc) Peer Firewall Peer TCP/IP HTTP Peer ID Firewall

13. 13 JXTA: Peer Groups Set of peers that share a common set of interests Specific management policy Peer group services Peer ID NetPeerGroup PeerGroupA PeerGroupB

14. 14 JXTA: Advertisements Every resource is described by an advertisement Peers Peer groups Communication channels Services … PeerGroup Advertisement: urn:jxta: uuid- BCBCDEABDBBBABEABBBABA000000 urn:jxta:uuid- BFEFDEDFBABAFRUDBACE00000001 My Group This group is to be used for my own testing

15. 15 JuxMem: an Architecture Proposal juxmem group cluster A group cluster B group cluster C group data group Physical architecture Logical architecture

16. 16 JuxMem API Alloc (size, attribs) Map (id, attribs) Put (id, value) Get (id) Lock (id) Unlock (id)

17. 17 Managing Memory Resources cluster group juxmem group Size: 8 MB Memory provided Provider advertisements: cluster group Cluster advertisements: juxmem group

18. 18 Allocation: How Does It Work? 2 1 3a 3b 4 5 6 8 MB?

19. 19 Managing Shared Data Blocks Allocate a memory block = create a data group Data blocks identified by the ID of the peer group Transparent access for clients via data ID Consistency Data blocks replicated on providers Simultaneous updates (logical multicast) Clients are not notified of updates Synchronization One lock per data block Other mechanisms: in progress

20. 20 Handling Peer Volatility Provider volatility A manager per peer group (cluster and data) Dynamic monitoring of available peers (cluster) Automatic replication of data blocks (data) Manager volatility Periodic exchange of heartbeats Dynamic replication of managers if needed on other peers

21. 21 Implementation and Preliminary Evaluation Implementation JXTA service, 5000 Java code lines Experimental setup PentiumII: 450 Mhz and 256 MB of RAM FastEthernet 100 Mb/s Linux 2.4 Number of nodes: 20 Experiment Study provider volatility

22. 22 Study: Provider Volatility (1) juxmem group cluster group data group Data size: one byte Replication degree = 3 Data manager not killed 1 client: 100 iterations lock-put-unlock 16 providers

23. 23 Study: Provider Volatility (2) juxmem group cluster group data group 1 client: 100 iterations lock-put-unlock 16 providers Data size: one byte Replication degree = 3 Data manager not killed

24. 24 Study: Provider Volatility (3) Internal locking during replication Guarantee consistency during replica creation Client is blocked juxmem group cluster group data group

25. 25 Study: Provider Volatility (4) JXTA/Java Expensive underlying JXTA-level dynamic channel management Reconfiguration time  11 seconds Targeted volatility is weaker ( >> 80 seconds)

26. 26 Conclusion A hierarchical architecture for a data sharing service for the grid Hybrid approach: DSM and P2P systems Transparent access to data blocks Persistent storage Mutable data: consistency guarantees Active support for peer volatility

27. 27 Ongoing Work Studies Replication strategies for fault tolerance Consistency protocols in a dynamic environment Co-scheduling computation and data distribution Manage data-data affinity Integrate high-speed networks: Myrinet, SCI. Goal: build a Grid Data Service GDS project: http://www.irisa.fr/GDS Extensive evaluation on realistic codes