Jonathan Stanton1Spring 2004 / Lecture 16Network IICS 184TCP Half-Pipe for ClustersDepartment of Computer ScienceGeorge Washington UniversityJonathan Stanton2Spring 2004 / Lecture 16• Half-pipe Anchoring: An Efficient Technique forMultiple Connection Handoff by Ravi Kokku,Ramakrishnan Rajamony, Lorenzo Alvisi, andHarrick Vin. International Conference on NetworkProtocols 2002.Additional ResourcesJonathan Stanton3Spring 2004 / Lecture 16Clusters for Network Servers• Now lets look at one “neat” low-level networktechnology and analyze it.• What is a cluster?– High Performance Cluster: Beowulf or MPP• Goal: Increase computation speed.– High Availability/Fault Tolerant Cluster: DB failover,LVM, Microsoft Cluster Server• Goal: Mask faults, increase service availability– High Throughput Cluster: Web farms, parallel DB• Goal: Increase service throughput• How do clients communicate with cluster?Jonathan Stanton4Spring 2004 / Lecture 16TCP for Clusters• Many network services are provided by clusters.– Web is the biggest example, but others include login,email, user directories, search engines, and applicationservers.– Clusters are increasingly heterogeneous and containspecialized components.• How does TCP interact with a cluster?– Can you connect one TCP session to a “cluster” instead ofa “host”?– Can you decouple the IP as identity and IP as address?– How can you minimize the bottleneck of one frontendapplication server?Jonathan Stanton5Spring 2004 / Lecture 16Multi-request TCP• One of the common uses of TCP is for client-servercommunication where the client sends requests andthe server sends data in reply.– Commonly, multiple requests are sent during the lifetimeof one TCP connection.– What if each request could best be served by differenthosts?Jonathan Stanton6Spring 2004 / Lecture 16Splitting TCPIdea: Split the TCP connection into two separate unidirectionalcomponents.Jonathan Stanton7Spring 2004 / Lecture 16Problem ModelJonathan Stanton8Spring 2004 / Lecture 16Existing Models• Frontend Model:– C1 does layer 7 switching, andforwards request along persistent C2or C3 to server.– Disadvantage: Scalability of frontendnode.• Backend Model:– C1 does layer 4 switching and round-robins the connections across theservers.– S1 forwards request to optimal serveralong persistent C2 or C3, and thenproxies the reply data.– Disadvantage: Overhead of transferthrough S1, does not supportheterogeneous clusters.Jonathan Stanton9Spring 2004 / Lecture 16Split Stack Architecture• DServ is the Designated server who acceptsincoming connections.– Is referred to as the Active TCP stack.• OServ is the Optimal server who actually servicesthe request of the client.– Is referred to as the Passive TCP stack.Jonathan Stanton10Spring 2004 / Lecture 16Handoff ProcedureTCP socket stateand requestUpdateSeq.endModifyDestination IPWorkCreate socket inEstablished stateAll packetsACKedClose ConnectionSuccess w/o FINDServ TCP State:OServ.IP,Client.IP, Client.port,Seq.start, Seq.endJonathan Stanton11Spring 2004 / Lecture 16Evaluation of Half-pipe• What new service does Half-pipe provide?• Does it increase performance? How?• When does it slow down performance?• How complex is it to deploy?Jonathan Stanton12Spring 2004 / Lecture 16Client Server1Client Server1Server2DirectSplit-stack0123455 10 15 20 25Connection lifetime (in ms)File size(in KB)DirectSplit-stack0.36400.72201.12104.08216.24LAN (0.2)%OverheadWAN delay(in ms)WAN effects (File size = 15KB)Half-Pipe PerformanceTables thanks to Ravi Kokku’s (UTA)presentation at ICNP 2002Jonathan Stanton13Spring 2004 / Lecture 16Performance for multi-requestsClientServer3Server1Server2Designatedserver024681012141 2 3 4 5 6Connection lifetime (in ms)Requests per connectionDirectSplit-stackSplit-stackTables thanks to Ravi Kokku’s (UTA)presentation at ICNP 2002Jonathan Stanton14Spring 2004 / Lecture 16Information Slide• Midterm Exam is tentatively scheduled to start ThursdayApril 1st, 2004. It will be a takehome midterm of roughly24 hours. Due at midnight Friday April 2nd.• Emulab: Swapin requires additional permissions. I will seewhat I can do about that, but in the meantime if it theexperiment is swapped out, email me and the TA and oneof us will activate it and send email to the class list to letyou know.• Lecture slides can be obtained at the course web