1 A BRIEF INTRODUCTION TO HIGH ASSURANCE CLOUD COMPUTING WITH ISIS2 Cornell University Ken Birman 2 About the Lecturer An introduction to the lecturer Ken Birman 3 Researcher in high assurance computing since joining Cornell in 1982 (PhD U.C. Berkeley). Currently Cornell’s N. Rama Rao Professor of Computer Science. ACM Fellow, Winner of IEEE Tsutomu Kanai Award Built the distributed software infrastructure used for a decade by the New York Stock Exchange, and still used in the French Air Traffic Control System, the US Navy AEGIS and several other mission-criticial systems. Contact information at http://www.cs.cornell.edu/ken 4 Segment I: The Cloud Landscape Introducing terminology Informal description of goals High Assurance and the Cloud 5 Cloud Computing: The new universal standard A technology for federating network services Easy to share data, deeply integrated with web pages Supports a wide range of media types But the cloud can’t offer high assurance today! A wave of sensitive applications is approaching (areas like mHealth, Smart power grid, eBanking, Smart cars...) They need strong guarantees... what can we do to help? How does today’s cloud work? 6 Client platform: browsers and “apps”, which are programs that exploit a stripped-down browser API Internet transports the data Data centers run “web services” that produce the pages we see, stream videos, etc Each step embodies weaknesses 7 The client system is vulnerable to loss of connectivity, compromise by downloaded code and infection by viruses and worms. The Internet layer is potentially unreliable The mapping of domain names to IP addresses is very complex (consequence of cloud need to “steer” traffic) Network reliability is much lower than it needs to be Much too easy to snoop on traffic or attack connections The Web Services infrastructure can fail or reconfigure abruptly, forcing the client to reconnect Recipe for high assurance 8 Design a system to fail only in “safe ways” Nobody gets hurt, but perhaps the system reports that it has gone offline Then do everything practical to enhance reliability, consistency, security, other needed properties Today: Focus on the web services running on the cloud data center Tradeoffs in “cloud space” 9 Often weak or lacking Required The properties we need are in tension! Snappy response: Every 100ms matters Elasticity: Load varies suddenly and dramatically, service replication levels need to vary accordingly Consistency: If distinct service replicas “talk” to multiple clients about something, they don’t say contradictory things. Fault-tolerance: If a replica crashes, the cloud self-heals Attack-tolerance: The service is very hard to attack. Security: Authenticated clients are limited to performing authorized actions in accordance with a policy Privacy: I can control who uses my data and how Today’s cloud: As fast as possible 10 In the race to offer the fastest possible services to the largest possible number of clients today’s cloud often gives up on other assurance properties In some sense the cloud is insecure and inconsistent by design! ... but does it have to be that way? Tomorrow: A high assurance cloud! 11 A single system needs to tell multiple kinds of assurance stories and not all in the same way An mHealth application: Needs to reassure the user that it is trustworthy Needs to help the developer make the right choices Must implement complex protocols correctly Must be a good citizen on the cloud data center 12 Segment II: Examples A few slides each on some challenging problems Each needs the cloud... but each needs some form of strong assurance guarantee too Example 1: Power grid 13 Today’s power grid has serious issues Wasteful: As much as 15% of power is lost just moving it around, and a great deal of “renewable” energy (solar, wind, tides) is lost because of poor integration with the standard grid Rigid: Ideally, the grid should “adapt” and move parcels of power much as the Internet moves packets. Dumb: even when it is obvious that we could optimize behavior, the grid uses old, inefficient techniques Goal: A “smart” power grid! How a small power grid operates 14 Power flows “like water” Path of least resistance Governed by Kirchoff’s Law Power enters at every generator, exits at every load Hierarchical structure: Primary “power busses” Secondary smaller local feeds 10-Generator, 39-bus New England System Technology to enable a smart grid 15 We’ll need to monitor power loads, frequency, current in real-time, reliably and securely Use this data to estimate the state of the grid and to predict its evolution over time Use those predictions to plan control actions: increase/decrease generation, borrow “reactive” power from neighboring regions, adapt pricing, etc Ultimately the grid will become a new kind of network. But must also be safe, efficient, and secure against both mishaps and even attack! Even mundane problems can hurt 16 California: Repeated episodes of market manipulation aimed at increasing profits for companies such as Enron that speculate on pricing Multi-state and multi-national rolling outages Causes turmoil for air traffic, ground traffic, telephone outages Will “smartness” also make grid more fragile? Risk of CyberAttacks? Control of the smart power grid 17 Suppose that a cloud control system speaks with “two voices” In physical infrastructure settings, consequences can be very costly “Canadian 50KV bus going offline” “Switch on the 50KV Canadian bus” Control of the smart power grid 18 Suppose that a cloud control system speaks with “two voices” In physical infrastructure settings, consequences can be very costly “Canadian 50KV bus going offline” Bang! “Switch on the 50KV Canadian bus” Power grid summary 19 To make it smart we need to monitor at a massive scale and use that to initiate control actions But for this to be safe, we need more that fast response and elasticity We also need security (so that attackers can’t take the grid down) ... and consistency (as we just saw) ... and fault-tolerance (since power systems often experience failures of various kinds) Example 2: mHealth 20 A term for everything outside the doctor’s office (but might be linked to electronic health records) Goal is to make your life better and healthier Encourage activity Discourage poor nutrician choices Help patients with chronic conditions manage their complex medical devices and medications Offer caregivers a window into health so that the patient can maintain independence What properties are needed in remote medical care systems? 21 Motion sensor, fall-detector Healthcare provider monitors large numbers of remote patients Medication station tracks, dispenses pills Integrated glucose monitor and Insulin pump receives instructions wirelessly Cloud Infrastructure Home healthcare application Durability... scalability... fast response 22 Mrs. Marsh has been dizzy. Her stomach is upset and she hasn’t been eating well, yet her blood sugars are high. Let’s stop the oral diabetes medication and increase her insulin, but we’ll need to monitor closely for a week Cloud Infrastructure Patient Records DB Need: Strong consistency and durability for data What do these terms mean? 23 Consistency: Even if accessed by multiple users concurrently, the data looks like a single database This sounds like it should obviously be true, but when the data is spread over multiple computers, if they don’t coordinate their actions, consistency can easily violated For example, perhaps machine 1 shows updates machine 2 never saw. Perhaps machine 3 sees all the updates but has the order confused. Each of these cases can cause serious inconsistencies. What do these terms mean? 24 Durability: Even if system components crash and then recover later, data will not be lost. Updates confuse things: before the update occurs, clearly it isn’t durable After the update is finished, it must have durable effect Question to pose: exactly when did it need to be durable? Usual answer: If the effect of an update survives a crash, then the update itself should also survive the crash Scalability 25 As we make the system larger, perforance remains good It needs to be able to support large numbers of clients and run on large numbers of cloud computing systems Fast response: Queries shouldn’t delay for long. Updates should have rapid effect on the data. Guarantees versus “best effort” 26 Today’s cloud systems work well in all of these ways but without providing strong guarantees except in certain very specialized cases, like Google’s new “Spanner” database Our challenge: can normal people who aren’t in the Google spanner development team also create trustworthy cloud computing solutions? mHealth summary 27 The needs of the system vary depending on what part of the system we focus on In our example, some aspects need durability in the sense of a logged database update, while others might accept durability through in-memory replication This illustrates one of many such tradeoffs If we had more time we could identify a number of additional issues of this kind How The Cloud Was Built 28 It is very hard to create software to run in cloud computing systems Everything must be automated You must follow many rules and use many packages So open source “tools” have become popular Examples: Hadoop (a version of MapReduce), Zookeeper, Graphlab, Pregel, Vowpal Wabbit, global file systems like GFS, etc. In this short class we will focus on process group tools and will use Isis2 as our main example. An obsession with speed... 29 At very large scale, either a thing is extremely fast, or unacceptably slow So everything we do must be shaped by speed! High assurance is not an option if the solution would be dramatically slower For example, the cloud computing community avoids databases. They founded the NoSQL movement (storage, but not as strong as a SQL database) for this reason. Similarly we must have speed in mind at all times! 30 Concept: Critical paths To understand speed, understand the limiting factors This forces us to think about critical paths What limits responsiveness? 31 Top priority: delay until a client receives a reply Critical path traces actions that contribute to this delay Update the monitoring and alarms criteria for Mrs. Marsh as follows… Service instance Response delay seen by end-user would include Internet latencies Service response delay Confirmed Critical path with complex services? 32 When we replicate information but want to be sure the data won’t be lost, critical path extends into the replicas Update the monitoring and alarms criteria for Mrs. Marsh as follows… Service instance Critical path Response delay seen by end-user would include Internet latencies Service response delay Critical path Confirmed Critical path Why do critical paths matter? 33 When we build complex systems it is hard to imagine how they will behave when we run them By thinking about the critical performance-limiting paths, we can focus our attention on specific elements and not think about the whole system By avoiding delays on the critical path, we bring benefits to the whole system! There are many critical applications 34 Cloud-hosted system to control transportation (think of Google’s smart cars) The cars have autonomy but they depend on data from the cloud and would have a much harder challenge if that data couldn’t be trusted Banking systems Today’s online banking systems are growing, but as they happens, more and more security issues arise Process control Chemical refineries, manufacturing plants, ... And they come with similar stories 35 In each case we can identify properties that are Absolutely needed for a cloud deployment Absolutely needed for safety And beyond that we might have other assurance properties that a particular use case doesn’t need The challenge will be to analyze each application, and then to translate its needs into cloud solutions 36 Segment III: Consistency We’ll drill down on the tradeoffs between durability and consistency Many cloud systems believe that consistency isn’t possible: CAP theorem Yet consistency underlies so many other guarantees Virtual synchrony model We’re going to drill down… 37 … on data and service replication Replication is at the center of cloud computing: With many replicas a service can handle many clients And those replicas need as much of the critical data to be local as possible So replication is a key technology. It even underlies security: we need to replicate the policy database and certificates that identify principals (clients, servers, etc) Consistency for replication 38 There are many ways to replicate information But it becomes tricky if the data or even the service evolves over time. Replication of changing data can leave a confusing mess if a request encounters stale versions. In some situations these errors can harm the client. In others, they could cause security violations. What do we mean by consistency? 39 A consistent distributed system will often have many components, but users observe behavior indistinguishable from that of a single-component reference system. Our power system example illustrated a form of inconsistency “Canadian 50KV bus going offline” Bang! “Switch on the 50KV Canadian bus” Theory of Consistency 40 There are some famous impossibility results Fischer, Lynch and Patterson: FLP theorem proves that any correct fault-tolerant protocol strong enough to solve “consensus” (a form of agreement) can also wedge in the event of certain sequences of failures. But those sequences turn out to be very rare. Brewer’s CAP theorem posits that you can only have two from {Consistency, Availability and Partition Tolerance}. But the proof holds only for a service running in a WAN, not for one in a single data center. Relate consistency to speed? 41 How costly is strong consistency? The cloud computing community debates this topic! It is a very contemporary question We usually pose the question in connection to replicating data. Strongly consistent data means “guaranteed to be correct and current”. Can cloud systems afford strong consistency? Weakly consistent data means “best effort but can have mistakes.” Facebook, eBay, Google all use weak consistency We will learn more about these topics 42 In today’s lecture we won’t “drill down” But in lecture 4 we will look more closely at these theoretical questions Mathematics is a valuable tool for cloud computing By making a correspondance of computing ideas to mathematics we can reason more rigorously Yet we will also find that some of the existing theory has limitations of its own 43 Segment IV: 2 Isis How does consistency look to the end user? What is it like to program with a powerful high assurance library like Isis2? 2 Isis System 44 A prebuilt technology that automates many of the hard tasks involved in replicating services and the data on which they depend Targets cloud computing settings Available in open-source from isis2.codeplex.com Intended to be easy to use… … but still at an early stage of development 2 Isis System 45 C# library (but callable from any .NET language) offering replication techniques for cloud computing developers Based on a model that fuses virtual synchrony and state machine replication models Research challenges center on creating protocols that function well despite cloud “events” Elasticity (sudden scale changes) Potentially heavily loads High node failure rates Concurrent (multithreaded) apps Long scheduling delays, resource contention Bursts of message loss Need for very rapid response times Community skeptical of “assurance properties” Isis2 makes developer’s life easier 46 Benefits of Using Formal model Formal model permits us to achieve correctness Isis2 is too complex to use formal methods as a development too, but does facilitate debugging (model checking) Think of Isis2 as a collection of modules, each with rigorously stated properties Importance of Sound Engineering Isis2 implementation needs to be fast, lean, easy to use Developer must see it as easier to use Isis2 than to build from scratch Seek great performance under “cloudy conditions” Forced to anticipate many styles of use 2 Isis makes developer’s life easier 47 Group g = new Group(“myGroup”); Dictionary<string,double> Values = new Dictionary<string,double>(); g.ViewHandlers += delegate(View v) { Console.Title = “myGroup members: “+v.members; }; g.Handlers[UPDATE] += delegate(string s, double v) { Values[s] = v; }; g.Handlers[LOOKUP] += delegate(string s) { g.Reply(Values[s]); }; g.Join(); First sets up group Join makes this entity a member. State transfer isn’t shown Then can multicast, query. Runtime callbacks to the “delegates” as events arrive Easy to request security (g.SetSecure), persistence g.Send(UPDATE, “Harry”, 20.75); List<double> resultlist = new List<double>(); nr = g.Query(ALL, LOOKUP, “Harry”, EOL, resultlist); “Consistency” model dictates the ordering aseen for event upcalls and the assumptions user can make 2 Isis makes developer’s life easier 48 Group g = new Group(“myGroup”); Dictionary<string,double> Values = new Dictionary<string,double>(); g.ViewHandlers += delegate(View v) { Console.Title = “myGroup members: “+v.members; }; g.Handlers[UPDATE] += delegate(string s, double v) { Values[s] = v; }; g.Handlers[LOOKUP] += delegate(string s) { g.Reply(Values[s]); }; g.Join(); First sets up group Join makes this entity a member. State transfer isn’t shown Then can multicast, query. Runtime callbacks to the “delegates” as events arrive Easy to request security (g.SetSecure), persistence g.Send(UPDATE, “Harry”, 20.75); List<double> resultlist = new List<double>(); nr = g.Query(ALL, LOOKUP, “Harry”, EOL, resultlist); “Consistency” model dictates the ordering seen for event upcalls and the assumptions user can make 2 Isis makes developer’s life easier 49 Group g = new Group(“myGroup”); Dictionary<string,double> Values = new Dictionary<string,double>(); g.ViewHandlers += delegate(View v) { Console.Title = “myGroup members: “+v.members; }; g.Handlers[UPDATE] += delegate(string s, double v) { Values[s] = v; }; g.Handlers[LOOKUP] += delegate(string s) { g.Reply(Values[s]); }; g.Join(); g.Send(UPDATE, “Harry”, 20.75); List<double> resultlist = new List<double>(); nr = g.Query(ALL, LOOKUP, “Harry”, EOL, resultlist); First sets up group Join makes this entity a member. State transfer isn’t shown Then can multicast, query. Runtime callbacks to the “delegates” as events arrive Easy to request security (g.SetSecure), persistence “Consistency” model dictates the ordering seen for event upcalls and the assumptions user can make 2 Isis makes developer’s life easier 50 Group g = new Group(“myGroup”); Dictionary<string,double> Values = new Dictionary<string,double>(); g.ViewHandlers += delegate(View v) { Console.Title = “myGroup members: “+v.members; }; g.Handlers[UPDATE] += delegate(string s, double v) { Values[s] = v; }; g.Handlers[LOOKUP] += delegate(string s) { g.Reply(Values[s]); }; g.Join(); First sets up group Join makes this entity a member. State transfer isn’t shown Then can multicast, query. Runtime callbacks to the “delegates” as events arrive Easy to request security (g.SetSecure), persistence g.Send(UPDATE, “Harry”, 20.75); List<double> resultlist = new List<double>(); nr = g.Query(ALL, LOOKUP, “Harry”, EOL, resultlist); “Consistency” model dictates the ordering seen for event upcalls and the assumptions user can make 2 Isis makes developer’s life easier 51 Group g = new Group(“myGroup”); Dictionary<string,double> Values = new Dictionary<string,double>(); g.ViewHandlers += delegate(View v) { Console.Title = “myGroup members: “+v.members; }; g.Handlers[UPDATE] += delegate(string s, double v) { Values[s] = v; }; g.Handlers[LOOKUP] += delegate(string s) { g.Reply(Values[s]); }; g.Join(); First sets up group Join makes this entity a member. State transfer isn’t shown Then can multicast, query. Runtime callbacks to the “delegates” as events arrive Easy to request security (g.SetSecure), persistence g.Send(UPDATE, “Harry”, 20.75); List<double> resultlist = new List<double>(); nr = g.Query(ALL, LOOKUP, “Harry”, EOL, resultlist); “Consistency” model dictates the ordering seen for event upcalls and the assumptions user can make Concept: A “multi-query” 52 Our lookup is Multicast to the group All members respond Lookup “Harry” in the Ithaca phone directory Front end A chance for parallelism Each can do part of the job: e.g. search 1/nth of a database Reduces response delays Names with Harry in them: .... With n replicas... ... we get an n times speedup! Our example was overly simple 53 it didn’t show the “state transfer” code Corresponds to the “white arrows” in time-line figure In Isis2 we have a way to make checkpoints State transfer: Some active member makes a checkpoint, and the joiner loads the state from it. The code looks like other operations in our example p q r s t Time: 0 10 20 30 40 50 60 70 Checkpoints can also be used to save group state during periods when all members are inactive Adding security: Just one line! 54 Group g = new Group(“myGroup”); First sets up group Dictionary<string,double> Values = new Dictionary<string,double>(); g.ViewHandlers += delegate(View v) { Console.Title = “myGroup members: “+v.members; }; g.Handlers[UPDATE] += delegate(string s, double v) { Values[s] = v; }; g.Handlers[LOOKUP] += delegate(string s) { g.Reply(Values[s]); }; g.SetSecure(myKey); g.Join(); g.Send(UPDATE, “Harry”, 20.75); List<double> resultlist = new List<double>(); nr = g.Query(ALL, LOOKUP, “Harry”, EOL, resultlist); Join makes this entity a member. State transfer isn’t shown Then can multicast, query. Runtime callbacks to the “delegates” as events arrive Easy to request security, persistence, tunnelling on TCP... “Consistency” model dictates the ordering seen for event upcalls and the assumptions user can make Some uses for process groups 55 To replicate data maintained by the members in memory To replicate actions taken on an external service such as a replicated database To ensure that all replicas are configured the same way To coordinate the processing of requests and load-balance To offer a way to parallelize processing by having each group member do part of the work Fault-tolerance via a backup scheme 2 Isis Summary 56 A library that you can invoke from a normal program written in a normal way It does the work of creating groups and sending multicasts and ensuring that the consistency model will be enforced The developer just tells it what to do. She thinks about a parallel distributed application. Virtual synchrony eliminates many hard problems Why not build it yourself from scratch? SafeSend and Send are two of the protocol components hosted over what we call the large-scale properties sandbox. The sandbox addresses issues like flow control, security, etc. All protocols share and benefit from those properties 57 Isis2 user object Isis2 user object Isis2 user object Other group members Membership Oracle The SandBox itself is mostly composed of “convergent” protocols that use probabilistic methods Isis2 library Send CausalSend OrderedSend SafeSend Query.... Flow Control Group instances and multicast protocols Group membership Reliable Sending Large Group Layer Fragmentation Dr. Multicast Sense Runtime Environment Message Library Group Security Platform Security Socket Mgt/Send/Rcv “Wrapped” locks TCP tunnels (overlay) Report suspected failures Views Oracle Membership Self-stabilizing Bootstrap Protocol Bounded Buffers These systems are complex, especially if you want to run on platforms like EC2 By using Isis2 you “inherit” 30 years of research on how to make it work Why focus on 2 Isis ? 58 This is a good question to ask In fact we could focus on any of a number of other technologies, including other multicast products Such as Spread, JGroups, C-Ensemble... But Isis2 is open source and specifically designed for cloud settings. (Also, Ken built it!) So since our class is short, we will look at Isis2 examples 59 Segment V: Performance Can Isis2 applications achieve the kinds of scalable performance and elasticity required in large cloud deployments? Revisit our notion of consistency 60 Let’s look again at our mHealth example We want the best possible performance but we also want to be sure that the application is “safe” for this kind of use We need consistency, yet also need snappy response and elasticity, especially in the monitoring component After all, it continuously monitors huge numbers of patients. What limits scalability? Speed of updates 61 Isis2 offers many ways to do updates RawSend, Send, CausalSend, OrderedSend, SafeSend Each has different consistency / durability guarantees As a developer, you’ll want to use the fastest option that is still safe in your setting ... Hence will need to understand how each works ... and how fast each solution will be Today we’ll just look at this superficially Example: Speed of updates 62 Isis2 offers several ways to do updates (we will visit them more carefully later) They have big performance implications But speed can have more than one definition! 2 Isis : Send v.s. in-memory SafeSend 63 Send scales best, but SafeSend with in-memory (rather than disk) logging and small numbers of acceptors isn’t terrible. Latency ops/second 64 Latency: Delay before external user sees action Ops/second: total throughput most purposes systems “like” Isis2 offer basic performance of about 1000 ops/second But by grouping requests into batches of ~50/request, services that can support ~50,000 ops/second are feasible Building them is challenging, but we won’t focus on that engineering topic in these lectures For Jitter: how “steady” are latencies? 65 The “spread” of latencies is much better (tighter) with Send: the 2-phase SafeSend protocol is sensitive to scheduling delays Cornell (Birman): No distribution restrictions. Flush delay as function of shard size 66 Flush is fairly fast if we only wait for acks from 3-5 members, but slow if we wait for all members. Isis2 lets developer set the threshold. Cornell (Birman): No distribution restrictions. So I want Send+Flush, right? 67 The problem is that the different solutions offer different guarantees The fastest solutions have weaker guarantees Using them safely involves understanding these properties in order to decide whether they are good enough for the desired purpose But there are subtle issues we don’t have time to discuss in today’s lecture. We will revisit tomorrow. Raw speed isn’t the whole story! 68 When building a system such as this we need to look at performance but also at steady behavior Here’s an example of a problem we ran into when doing the experiments I just showed you As we’ll see, Isis2 had an instability. We think we’ve fixed it but it illustrates an important point The experiment we did 69 We made a timeline picture from left to right One node (the bottom one) sends multicasts The others log the time of receipt We graphed the delay, sorted from slowest (top) to fastest (bottom) delays Here’s what we saw Debugging: Stabilization bug 70 Birman: DARPA MRC Kickoff, Washington, Nov 3-4 2011 As the application ran, it slowed down! 71 At first the system was fast: even the slowest nodes at the top had short delays But within a few multicasts they slowed down Then something “resets” them and they speed up We tracked it down to a problem with garbage collection in our system Modifying that protocol helped smooth things out Debugging : Stabilization bug fixed 72 Birman: DARPA MRC Kickoff, Washington, Nov 3-4 2011 Debugging : 358-node run slowdown 358-node run slowdown: Zoom in 358-node run slowdown: Filter Summary of insights from example? 76 Tools like Isis2 enable us to build cloud-scale replication based services with strong guarantees But today, at least, they demand a lot from the developer, who needs to really understand the choices and their implications As Isis2 evolves, this problem will be reduced: the system will eventually automate many decisions, including picking the right update primitives for you 77 Segment V: Conclusions We’ve scratched the surface but there is much more to be explored Cornell’s high assurance researchers are creating solutions for tomorrow’s demanding applications Key take-away points 78 Cloud computing, today, isn’t very friendly to high assurance applications This is a problem because those applications are increasingly forced to migrate to the cloud for reasons of cost, scalability or just because the cloud is the dominant paradigm today But we can already use tools like Isis2 to solve these problems and as they become easier to work with, the community able to build these solutions will grow Key take-away points 79 With Isis2 we can easily create programs that run on cloud platforms like EC2 or even Android mobile They form into groups and coordinate or replicate data or actions via group primitives The concept is powerful and easily visualized But tuning and doing sophisticated fault-tolerance remains challenging. In the remaining lectures we will explore these issues The last word... 80 The word on the street is that cloud computing will rule but that the cloud can’t do high assurance But the word in the hallways at Cornell differs! see Isis2 as our proof-by-demonstration that it can be done Even so, the engineering challenge remains enormous We Learning more 81 Stay in the class. We’ll show you how! Download the Isis2 system from isis2.codeplex.com You can access the user’s manual The code itself (currently v2.xxx, a very stable release) And we maintain a discussion and issues board there Learning more 82 My textbook covers this topic in depth “Guide to Reliable Distributed Systems: Building HighAssurance Applications and Cloud-Hosted Services” Ken Birman. Springer Verlag, February 2012 A paper focused entirely on today’s topic is: Overcoming CAP with Consistent Soft-State Replication. Kenneth P. Birman, D. Freedman, Q. Huang and Patrick Dowell. IEEE Computer Magazine (special issue on “The Growing Impact of the CAP Theorem”). Volume 12. pp. 50-58. February 2012. You can download a copy from: http://www.cs.cornell.edu/projects/quicksilver/pubs.html