Quality of Service (QoS) Enabled Dissemination of Managed Information Objects in a Publish-Subscribe-Query

Transcription

1 Quality of Service (QoS) Enabled Dissemination of Managed Information Objects in a Publish-Subscribe-Query Information Broker Dr. Joe Loyall BBN Technologies The Boeing Company Florida Institute for Human Machine Cognition Vanderbilt University it Institute t for Software Integrated t Systems April 16,

2 Core Concepts of Pub-Sub-Query Information Management Consumers Access control & web-based b administration Subscription predicate Standardized connections for API clients (users of the information management system) Control Publication to make information available Subscribe to request future information Query to request past information Requests are made using predicates over information attributes (metadata) Using a rich, standards-based predicate language, XPath & Xquery Information is encapsulated into an organized MIO structure Metadata describing the information to facilitate brokering Payload, the original data Producers MIO Common Access <metadata> <baseobject> <InfoObjectType> <Name>Image</Name> </InfoObjectType> <Publisher>Intel COI </Publisher> <Latitude> </Latitude> <Longitude> </Longitude> </baseobject> </metadata> Archive Broker Match Data store Query predicate Managed Information Object (MIO) 2

3 Quality of Service is a Crucial Aspect of Information Management for Net-Centric Operations QoS management for net-centric information systems strives to provide a predictable high level of mission effectiveness & user satisfaction within available resources Providing timely delivery of Responding rapidly to priority information needed by tactical shifts and unfolding situations users in mobile scenarios Operating in a manner that is Tailoring and prioritizing robust to failures and information based on mission intermittent communications needs and importance 3

4 QoS Properties of Information Timeliness The speed at which brokering happens The latency through the information brokering system Completeness The number of useful data objects delivered compared to what are available to be delivered Fidelity The preservation of information content Accuracy The correctness of delivered information Smoothness The predictability of performance & lack of jitter, i.e., low variance Suitability How well a response matches a user s needs 4 Network & platform level QoS affect data traffic only and do not address the QoS properties of information

5 Core Information Management Services 5 Sends outgoing MIOs Matches MIOs & registered subscriber predicates Subscribers Producers Dissemination Service Query results Query clients Register Subscription Query predicate Registered Predicates Submission Service Broker Archive Service Published MIO Query results Processes queries & getnext calls Query Service Archive Service MIO Data store = Receives incoming MIOs Inserts MIOs into repository Managed Information Object (MIO)

6 Blocked outgoing traffic (timeliness) Dropped outgoing traffic (completeness) MIO size & variation (timeliness, smoothness) Predicate complexity & set size (timeliness, smoothness) Predicate expressiveness ess e ess (completeness, suitability) 6 Effect of IM Services on QoS Subscribers Dissemination Service Query results Query clients Register Subscription Query predicate Registered Predicates Producers Submission Service Broker Archive Service Published MIO Query results Query complexity, size & indexing of DB (timeliness, smoothness) Query expressiveness (completeness, suitability) Query Service Archive Service MIO Data store Blocked clients (timeliness) Dropped MIOs (completeness) Queued MIOs (timeliness, smoothness) Stored MIOs (timeliness) MIOs not archived (completeness) Execution time of all services affect timeliness Variations in execution time affect smoothness Nothing explicitly affects fidelity or accuracy

7 QED Architecture Publishing Client Subscribing Client Querying Client Host boundaries MIO shaping Configurable transport Diff. queues LQM controller QoS monitor Configurable transport Diff. queues LQM controller QoS monitor Configurable transport Diff. queues LQM controller QoS monitor Configurable transport (rcvg end) Configurable transport (sending end) Differentiated queues ISQM Differentiated queues LQM controller MIO shaping Predicate Evaluator MIO archive LQM ctlr LQM controller LQM controller QoS monitor Web Administration QoS Administration 7 Apollo v1.0 QoS administration Allows entry of QoS policies & mission information Information space QoS manager Aggregates QoS policies Local QoS managers Enforce QoS policies Monitor & maintain local QoS behaviors Differentiated queues In transport, for priority dissemination In core, for priority MIO processing MIO shaping Reduce demands on resources Configurable transport Options for reliable, ordered, unreliable, unordered combinations QoS monitors Provide visibility for QoS managers & users

8 QED Administration Service Supports selection & application of policies Based on filters defined over properties of actors, information, or operations Assignment of relative importance Assignment of client preferences Creation of new filters Grouping based on mission, i role, etc. 8

9 QoS Policies in QED Client Sequence creation Task Mgt LQM QED Admin ClientPolicy objects ISQM Service Seq. Properties Policies define a mapping from a context C to an importance (i) & set of QoS preferences (P): C (i, P) Context is the range of functions Fairness over observable properties of Actors strategy (A), Operations (O), Information (M), Policy actions Policy Mgt Svc Resources (R) Policy, Seq. ID C = f(a) g(o) h(m) j(r) Policies are maintained in the Policy Fairness strategy Management service Sequence attributes The ISQM distributes policies & fairness strategies to Local QoS Dissem. Svc LQM Managers We currently have two LQMs Task management LQM manages CPU intensive tasks, i.e., brokering, query, getnextresultset, & archive Dissemination service LQM manages bandwidth intensive tasks 9

10 QED Approach to Managing CPU Contention Using Priority Bins & Task Scheduling Task queues Thread Pool Registered Predicates Broker task Producer Publish MIO Broker task Archive task Insert task Extract task Get thread to assign to task Dissemination Worker Predicate Evaluation Query Query client Task Creation Task Mgt LQM Archive task Process boundary GetNext- ResultSet Query task GetNext task Archive Worker Query task GetNext task Query Worker Archive/ Query Service MIO Data store CPU intensive operations (brokering, queries, archival) are enqueued as tasks Use a configurable number of bins (queues) Each bin defined by importance (operation, MIO, client) & cost (expected duration of task) Bins allow for linear time insertion (vs. logn insertion time for priority queue) Number of bins determine granularity of control (one bin per task = priority queue) Assign threads to tasks in order of priority Policies define scheduling (thread assignment) strategies Weighted Fair Queuing (WFQ) Each queue i is serviced by a number of threads associated with its relative weight Strict Queuing available threads always serve highest priority tasks & tasks can be starved 10

11 QED Approach to Managing Bandwidth Contention Bandwidth Mgr Client queues Mocket Client Broker Task Client queues Set(subs), Dissemination MIO Mocket Client Svc Dissemination task Client queues Mocket Client 11 Dissemination i service creates When bandwidth is available, BW manager dissemination tasks semaphore unblocks the next diss. task Each diss. task is placed on a queue for its BW manager calculates the time to send destination s mocket based on its Size of MIO/amount of bandwidth importance Each client queue sends diss. tasks in priority order to BW manager Block on a getpermissiontosend semaphore MIO is sent through h the appropriate mocket BW manager blocks while MIO is in transit Time to send Each mocket should not have MIOs in queues

12 QoS Monitoring in QED Challenges for QoS Monitoring A common mechanism for metric registration, update & lookup. Mediate differences in rate, units, format, etc. Maintain different levels of granularity & abstraction to be used by different services. Provide seamless (and efficient) mechanism for monitoring information sharing between client & server platforms. QED Approach A modular architecture allows for any component to register metrics of interest. Metrics are stored as time-series, providing: Last value, average, variance, trend All information is hierarchically organized by the XLayer substrate so it can be easily retrieved. Monitoring information is shared automatically between nodes as part of the route advertisement messages (or through a local message dissemination service) Access the metrics OtherService 1 OtherService 2 Monitor Provides data to compute metrics Monitoring Service Monitoring Component A Monitoring Component B Monitoring Component C Metric Recorder Other components or applications Node 1 Node 2 Java Monitoring Service Component a Push metrics to Retrieve the XLayer Component b Metrics Proxy Java Monitoring Service Component a Push metrics to the XLayer Retrieve Component b Metrics Proxy Other components or applications out XLayerProxy in Metrics Host, Net data Link Info structure XLayer Host, Net Metrics Link Info data structure XLayer XLayer Metrics data structure Network 12

13 CPU Bound Example Differentiated Priorities with Weighted Fair Queuing Policy Weighted policy services higher priorities more frequently without starving any priority Type1 workflow is higher importance (medium) than Type2 (low) MioType1 objects are getting through at a higher rate & with low & steady latency Type1 (medium importance) is green; Type2 (low importance) is blue Relatively high rate & low latency Lower performance, but no starvation MioType2 objects are eventually getting through Weights are different for low, medium, & high importance 13

14 Network Bound Example Differentiated Priorities with Weighted Fair Queuing Policy Mockets currently implements a form of weighted fair queuing (items on queues are promoted rather than being starved) MioType1 objects have low & steady latency MioType2 objects are getting g through but at increasing latency All situations of increasing latency (in this & other demo graphs) will top out at a steady latency when queues get full & publishers block 14 MIO Type 1 (green, medium importance) has relatively low latency MIO Type 2 (blue, low importance) has relatively higher latency, but no starvation

15 Summary of Rates & Latencies in the AFRL Demonstration Scenarios Mean Receive Rate (Hz) Mean Latency (sec) Base Equal Weighted Strict Base Equal Weighted Strict CPU Bound (Queues) Type Type > 100 Infinite Network Type > Bound N/A (Mockets) Type > N/A Overhead Low overhead of QED Overall improvement comparison Replacement of Queue configuration under load existing services throttles clients to Additional services, resources e.g., ISQM are not in the Mockets more efficient runtime path than JMS 15 Green = MEDIUM importance Blue = LOW importance Black = undifferentiated service

16 Summary of Rates & Latencies in the AFRL Demonstration Scenarios Mean Receive Rate (Hz) Mean Latency (sec) Base Equal Weighted Strict Base Equal Weighted Strict CPU Bound (Queues) Type Type > 100 Infinite Network Type > Bound N/A (Mockets) Type > N/A Improved control over information QoS Significantly higher QoS provided to higher importance information Higher throughput, 2x over baseline Lower latency than baseline and low importance MIOs 16 Green = MEDIUM importance Blue = LOW importance Black = undifferentiated service

17 Concluding Remarks QED is developing QoS PSQ systems present challenges for management services for PSQ QoS management information management services Anonymous pub-sub makes the Mission-level QoS policies and QoS administration destination unknown at information entry into system Significant dynamism in resource usage, e.g., QoS policies and aggregate QoS management Local QoS managers, including task and bandwidth management QoS mechanisms, including differentiated queuing, predicate partitioning, i and information shaping QoS monitoring, visualization, and feedback Managed transport The current QED prototype has shown an improvement over the IM baseline, with little overhead (so far) Current work is ongoing MIO sizes, Time of predicate processing Query result set size Fanout of publication to multiple subscribers R&D challenges remain Intuitive capture of mission-level abstractions Optimization heuristics for aggregate QoS in PSQ systems Client QoS preferences Disruption tolerance 17