Transcript unreliable
(Yet another) Replicated Web Server Alberto Bartoli Bologna Austro-Hungarian Empire Italy Germany Yugoslavia Allied Administration Italy 1918 June 1945 1954 Previous work in the area Enriched Virtual Synchrony ( group communication for on-line state transfer) “old stuff”; IEEE ToC 1997; with Ozalp On-line reconfiguration of replicated databases based on group communication “algorithm”; DSN 2001; with Bettina and Ozalp Probably not very useful / applicable for Adapt Main interests / On-going work Service replication with sequential consistency in unreliable networks “algorithm”; unpublished; with Ozalp Implemented in a replicated Tomcat in Trieste Group communication support for replicated databases “algorithm”; unpublished; with Bettina Hope they will be useful for Adapt ! Service Replication: Scenario Unreliable Network C S C connected with only one server at a time Reads executed locally Writes executed at several servers (without C being involved) If C does not receive a response (e.g., TCP connection breaks), C may send the very same request to any server Unreliable Network Service Replication: Guarantees Unreliable Network C S “Sequential Consistency” Each write is executed only once Unreliable Each read issued by C reflects all previous writes issued by C Network Recall: Even across different connections to possibly different servers Really Useful ? Connected with one server at a time Can retransmit to any server until seeing a response Always sees the effect of previous requests Exactly-once Useful for users that navigate the web Crucial for integration of remote applications (if we want them to be reliable) Web services Adapt A note on existing solutions They do not provide those guarantees automatically 1) The service is “reliable only if you are lucky”, or, 2) Programmers must implement suitable support My ultra-strong belief: We will have unreliable services Even services supposed to be reliable (see my FuDiCo talk: http://www.cs.unibo.it/fudico/) Status (I) Trieste has a working prototype of this stuff: Tomcat with in-memory session replication Client: Browser Connected with one server at a time Can retransmit to any server until seeing a response Always sees the effect of previous requests Exactly-once Server: “normal servlets” Each session object is replicated “Reads”: On the local replica (if up-to-date, otherwise HTTP redirect) “Writes”: On all replicas in the primary partition Reintegration (state transfer): Blind On-line and negotiated Status (II) Classes of distributed web servers: 1) Each server has its own IP address 2) Each server has its own IP address, but this is not visible to the outside (load-balancing switch in front: clients only see this IP address) 3) All servers share the same IP address We use 1, but it should work also with 2 Substantial redesign required for 3 Load balancing (for 1) Mechanism: Each server can communicate a “load index” to all other servers (quick, very efficient) Policy: Not yet investigated Status (III) Tomcat Session Replication Load Balancing “Nicer” GC Interface Group Communication Developed at Johns Hopkins JMiramare JBora Spread GC for RepDB (I) “Improvement” of GC interface suitable for replicated databases Key issues: Site reintegration cannot occur instantaneously The traditional notion of primary partition is not sufficient (building a totally-ordered sequence of views does not allow you to recover from a total failure) Useful for Adapt ? Probably its “dynamic voting portion”... GC for RepDB (II) Tomcat Signal whether a view is Primary Majority-based choice (not very Adapt -able) JMiramare JBora We should base it on dynamic voting (much more Adapt -able) Spread Human Resources Available / Constraints 1 programmer (hired Sept. 1st) Maybe a PhD student starting next January (maybe) Master-level students Computer Science in Trieste: Full professors: Associate professors: Research associates: PhD students: EEUU-funded projects in CS: 1 “almost retired” 1 “computer artist”, 1 has never done any research 1 very good... in signal processing and robotics 1 in robotics 0 Related “Projects” Collaboration with a research group in Pisa (CNUCE-CNR): Very good at performance analysis by simulation A PhD student has been working for 2 years on QoS-based load balancing Geographically replicated, read-only web servers Next year he should work on load balancing for “JMiramare” Useful for Adapt ? If we obtain interesting results, they could be additional selling points What we can contribute Tomcat JMiramare Replacement / Extension to JMiramare for stateful web services ? JBora Session passivation ? Partial replication ? Geographical replication ? Messaging layer ? Spread