A Survey on Parallel Computing in Heterogeneous Grid Environments

1. A Survey on Parallel Computing in Heterogeneous Grid Environments Takeshi Sekiya Chikayama-Taura Laboratory M1 Nov 24, 2006

2. Dynamic Change of CPU/Network Load Parallel Computingin Grid Environments • Increase opportunity in which we can use multi cluster environments • But, schemes for stand alone clusters cause problems in grid-like usage • New mechanisms are needed • Handling heterogeneity • Firewall/NAT traversal • Adaptation to dynamic environment • Monitoring Heterogeneous hardware and software Firewall/ NAT Maintenance Complex Configuration Failure Difficult to Know What’s Happening

3. Heterogeneous Environments • Heterogeneous machines • Binaries are different • Complex configuration are required when hardware/software is different • Heterogeneous networks • Overheads of synchronization in parallel application with different latency/bandwidth • Firewalls/NATs

4. Firewall/NAT • Firewalls/NATs hinder bi-directional connectivity • Bi-directional TCP/IP connectivity needs to be provided to support a wide spectrum of applications Firewall or NAT

5. Solutions to the Internet Asymmetric-Connectivity Problem • MPI Environment on Grid with Virtual Machines [Tachibana et al. 2006] • Xen for VM and VPN for Virtual Network • Low cost VM migration • ViNe [Tsugawa et al. 2006] • A host named Virtual Router • Overlay network base • WOW [Ganguly et al. 2006]

6. Outline • Introduction • WOW • IPOP: IP over P2P • Routing IP on the P2P Overlay • Connection Setup • Joining an Existing Network • NAT Traversal • Experiments • Summary

7. Objective and Approach • The system architected to … • Adapt heterogeneous environments • Present to end-users a cluster-like environment • Scale to large number of nodes • Facilitate the addition of nodes through self-organization of virtual network • Less manual configuration • Approach with Virtualization • Virtual Machines • Homogeneous software • Self-organizing overlay network • All-to-all connectivity

8. A homogeneous software environment Offering opportunities for load balancing and fault tolerance Users can use pre-configured systems Linux distribution Libraries and softwares Virtual Machine

9. NAT Virtual Network Virtual Grid Cluster IPOP (IP over P2P) P2P Network Physical Infrastructure firewall P2P overlay network

10. IPOP [Ganguly et al. 2006] • Characteristics • A virtual IP address space • Self-organizing • Architecture • IP tunneling over P2P • A virtualized network interface (tap) captures virtual IP packets • Brunet P2P overlay network

11. Ethernet Frame Ethernet Frame IP Packet IP Packet Brunet Message IP Packet Capturing Virtual IP Packets • The tap appears as a network interface from applications • IPOP translates virtual IP addresses to Brunet P2P network addresses application Tunneling application tap tap IPOP IPOP

12. Ring-structured overlay Organized connections Near: with neighbors Far: across the ring 160 bit SHA-1 hash address Greedy routing Each node has constant number of connections O(log2(n)) overlay hops Brunet P2P n4 n3 n5 n2 n6 Multi hop path from n1 to n7 n1 n7 n8 n12 n9 n11 n10

13. Node A wishes to connect to node B A sends a CTM (Connect To Me) request to B over P2P network The CTM request contains A’s URI When B receives the CTM request, B sends a CTM reply to A The CTM reply contains B’s URI Connection SetupConnection Protocol CTM reply CTM request A B URI (Uniform Resource Indicator) ex.) brunet.tcp:192.0.0.1:1024

14. B sends a link request message to A over the physical network When A receives the link request, A simply responds with a link reply message Finally, new connection is established between A and B Connection SetupLinking Protocol Direct connection A to B link request A B link reply

15. Race condition may occur because linking protocol is initiated by both peers Linking Race Condition (1) link request link request link reply link reply Both attempts succeed

16. Check no existing connection or connection attempt, when nodes receive link request When nodes receive link error, they restart protocol with random back-off Linking Race Condition (2) link request link request Active linking on? link error link error link request Random back-off link reply

17. A new node N creates a leaf connection to an initial node I by directly using linking protocol I acts as forwarding agent for N Joining an Existing NetworkLeaf Connection Correct position of new node Initial node I Leaf connection New node N

18. N sends a CTM request addressed to itself over P2P network the CTM request contains N’s URI A CTM request is received by right and left neighbors, since N is still not in the ring Joining an Existing NetworkSend CTM request Left neighbor L CTM request Right neighbor R Initial node I New node N

19. L and R send CTM reply including their URI to I I forwards CTM reply to N Joining an Existing NetworkSend CTM reply Left neighbor L CTM reply Right neighbor R Initial node I CTM reply New node N

20. Start linking protocol L and R send link request message to N over the physical network Joining an Existing NetworkLinking Protocol Left neighbor L Link request Right neighbor R Initial node I Link request New node N

21. N forms connections with neighbors and is in ring Acquires “far” connections Joining an Existing NetworkComplete Joining Left neighbor L New node N Right neighbor R Initial node I

22. Adaptive Shortcut Creation • High latencies were observed in experiments due to multi-hop overlay routing • Shortcut creation • Count IPOP packets to other nodes • When number of packets within an interval exceeds threshold, initiate connection setup • Because overhead incurred during maintenance connections, drop connections no longer in use

23. NAT IP: 192.168.0.2 IP: 133.11.238.100 IP: 157.82.13.244 Host b Host a NAT Src: 192.168.0.2:5000 Dst: 157.82.13.244:80 Src: 133.11.23.100:6000 Dst: 157.82.13.244:80 Src: 157.82.13.244:80 Dst: 192.168.0.2:5000 Src: 157.82.13.244:80 Dst: 133.11.23.100:6000 Private Network Global Network NAT Table 192.168.0.2:5000 ⇔ 133.11.23.100:6000

24. NAT TraversalUDP Hole Punching IP: A IP: N IP: M IP: B Host A NAT NAT Host B Src: A:a Dst: M:m Src: N:n Dst: M:m Src: M:m Dst: A:a Src: M:m Dst: N:n Src: B:b Dst: N:n NAT Table A:a ⇔ N:n NAT Table M:m ⇔ B:b

25. Experimental Setup Hosts: 2.0 GHz Xeon, Linux 2.4.20, VMware GSX Hosts: 2.4GHz Xeon, Linux 2.4.20, VMware GSX Host: 1.3GHz P-III Linux 2.4.21 VMPlayer Host: 1.7GHz P4, Win XP SP2, VMPlayer 34 compute nodes, 118 P2P router nodes on PlanetLab

26. Node A: IPOP node Node B: new joining node A and B are in different network domains with NAT B sends ICMP packets to A at 1sec intervals Within period 1 (about 3 seconds), B establish a route to other nodes Within period 2 (about 28seconds), B establish a shortcut connections to A Experiment 1Joining and Shortcut Connections

27. Parallelization with PVM based master-workers model FastDNAml has a high computation-to-communication ratio Dynamic task assignment tolerates performance heterogeneities among computing nodes Experiment 2PVM parallel application: FastDNAml (1) Master Task Pool Workers

28. Experiment 2PVM parallel application: FastDNAml (2) • The execution with shortcuts enabled is 24% faster than that with shortcuts disabled • The parallel speedup is 13.6x • 23x is reported in previous work in homogeneous cluster

29. Summary • Introduced WOW • Scalable, fault-resilient and low management infrastructure • Future works • Research on middleware which is easy to use for heterogeneous adaptive Grid environment