Research Challenges of Autonomic Computing. Jeffrey O.Kephart, IBM Thomas J. Watson Research Center, 19Skyline Drive, Hawthorne, NY 10532, USAPresenter: Djuradj Babic Agnostic: Juan Carlos Martinez1. - Difference from Autonomic Computing and a system with a lot of validations init.Autonomic Systems are systems that contain validations also. The main differencebetween an Autonomic System and a System with a lot of validations is that anAutonomic system takes actions not only due to Anticipated Events but also due toUnanticipated Events.2. - In order for Autonomic elements to interact, they should follow a standard. Somy question is there an organization that is in charge of defining this standard?There is not a specific organization in charge of defining a standard for it. There are someattempts to create a standard format for autonomic element interoperation such CommonBase Event Format developed by OASIS Web Services Distributed ManagementTechnical Community.For more information about the available standards for Autonomic Computing we can seethe following chart gotten from a paper from IBM called “An architectural blueprint for autonomic computing”.3. - Ideally, a good portion of the rule authoring should be based on some form ofmachine learning. What techniques different from Artificial Intelligence could weuse to achieve this?Data Mining could be used for extracting rules and pattern out of massive data sets.4. - How can we compensate the loss in performance we get using AutonomicComputing? Could the Self-Optimization feature help? If so, how?The loss in performance due to the feature of adaptability inside Autonomic Systemsmight in the long run be compensated because classic systems will consume aconsiderable time for repetitive tasks or rebooting applications whereas an autonomicsystem will adapt itself to handle this kind of situations. 5.- In the paper, it is suggested that we should look for the remedy of a problem bylooking at its symptoms, but what if there is more than one problem with the samesymptom?We would deal trying to narrow the causes of that symptom as much as possible and tryto see which previous actions (remedies) from the past were taken to solve this kind ofproblem keeping into the account also the known results from these actions. However,this will not guarantee that we might eventually fall into prescribing a wrong remedy.6. - Is it always better to look for a remedy instead of just rebooting the application?If not, when for example?Even though the remedy should be seen as the best approach for solving a problem, thereare some isolated instances where rebooting might be more appropriate due to facts likefor example no data loss, time performance, etc. For this approach we have the RecursiveMicro Rebooting where individual system components are designed as crash onlymeaning that they are basically designed to reboot if something goes wrong. They aredesigned in such a way that rebooting them usually takes 1/10 seconds.7. - Priorities. Which feature should be first considered in an autonomic system:Self-Healing, Self-Managing, Self-Protection, Self-Configuration, or Self-Configuring?All 4 are important. If one of them is disregarded some problems may occur with theother “Self” aspects due to the fact that they are self complimentary. However, prioritiesmight and most of the the times will depend on the policies given by the system.8. - Conflicts between Self-Configuring and Self-Protection? For example one of theSelf-Configuring actions is to open Port XXX and another feature from Self-Protection closes it? Which action should be taken and why?Depends on the high level policies established by the system administrator9. - What would be the best system storage system for the knowledge acquired bythe machine learning?We cannot say that a “best” storage exists. Databases would be a nice option, however,the best thing here would be to have a customized database in such a way that a bigportion of main memory would be in charge of storing most crucial and important datarelevant to functionality of the Autonomic system. If we are talking about products on theout on the market, data-warehouses would be a good solution because they are good withhandling a big amount of data that will be acquired by machine learning feature from theAutonomic System.10. - There are benchmarks to measure different features of systems; however, thepaper mentions lack of benchmarks for the aspects of Self-Configuration, Self-Healing and Self-Protection. What kind of approach could we follow to have suchmeasurements?Basically the approach would come from the side of dependabilities. Attempts to reachthis goal are Dependability benchmarks and Incipient Autonomic benchmarks for self-configuring and self-healing.Final CommentsThe concept of “Autonomic Systems” is not something that is completely new nowadays.There have been some kind of autonomic features in the past in some technologies orsystems, for example, the autonomic way in which routers learn which path shouldfollow in order to send a packet of data.About testing in autonomic systems, it is much harder because for every test we want toknow we would have to have very clear what the precondition and postcondition are andsince in an autonomic system there might be changes at runtime, it would be impossibleto predict what the current status of the system is at any moment.The author of this paper is a more AI oriented guy and that's why for him MachineLearning is crucial for a system to be considered autonomic; however, that is not acompulsory aspect. In other words, an autonomic system might be built without havingmachine