Accommodating Bursts in Distributed Stream Processing Systems Yannis Drougas, ESRI Vana Kalogeraki, AUEB

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 2 Large class of emerging applications in which data streams must be processed online Example applications include: –Traffic Monitoring –Sensor network data processing –Stock Exchange data filtering –Surveillance –Network monitoring Stream Processing Applications

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 3 Distributed Stream Processing Systems High-volume, continuous input streams Processed result streams On-line processing functions / continuous query operators implemented on each node: ClusteringCorrelationFilteringAggregation...

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 4 Data is produced continuously, in large volumes and at high rates Data has to be processed in a timely manner, e.g. within a deadline Application input rates fluctuate notably and abruptly E.g.: –increasing traffic volume due to a DoS attack –large number of messages generated when there is a time-critical event in the sensor network field It is important that no data is lost and that time-critical events are processed and delivered in a timely manner Stream Processing Applications Characteristics Clustering Filtering Join

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 5 Previous Work The majority of previous work 2003, VLDB2007, 2006, 2005] has focused on optimizing a given utility function –Some solutions employ data admission –Others consider the optimal placement of tasks on nodes The case where load patterns can be predicted has also been studied ICDE 2005] QoS management to predict query workload Common ways to deal with an overload: Admission control or resource reservation techniques [Abeni & RTSS’04] can result in under-utilization QoS degradation is a reactive technique and occurs only after burst has occurred IPDPS’05]

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 6 Our Problem We focus on the problem of addressing bursts of input data rate –Can we accommodate a burst without having to reserve resources a priori? –How should we react to a burst? Benefits: –Lost data units due to bursts are minimized –No QoS degradation or data admission –No under-utilization, dynamic reservation used Challenges: –Highly dynamic / unpredictable environment –Multiple limiting resource types –Plan must be applied in real-time for the burst

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 7 Roadmap Motivation and Background System Architecture Burst Handling Mechanism –Feasible Region –Pareto Points –Online burst management through rate assignment Experimental Evaluation Conclusion

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 8 System Architecture Overlay network consisted of processing nodes. Built over a DHT (currently, Pastry). The physical (IP) network. Application layer: Execution of stream processing applications.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems Synergy Node Architecture Replica Placement places components Load Balancing and Load Prediction detect hot-spots Migration Engine alleviates hot-spots Application Composition and QoS Projection instantiate applications Execution Scheduler provides real-time execution Node Monitor measures processor and bandwidth Discovery locates streams and components Routing transfers streaming data 9

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 10 Distributed Stream Processing Application A stream processing application is represented as a graph. The application is executed collaboratively by peers of the system that invoke the appropriate services. Multiple applications can run simultaneously sharing the system resources. A service can be instantiated on more than one nodes. A service instantiation on a node is a component.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 11 Each component c i is characterized by its resource requirements: The selectivity of the component c i is given by: Each node n is characterized by the availability of its resources: System Model

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 12 Optimization Problem Capacity Constraints: Flow Conservation Constraints: We need to come up with a plan that satisfies the above constraints and minimizes the likelihood of missing data due to bursts. The plan should determine the rate assignment of each component r c i

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 13 Feasible Region Assume we have Q applications The state of the system at any given time can be described by a point in the Q-dimensional space The feasible region is the set of all points (application input rates combinations) that nodes in the given distributed stream processing system can accommodate without any data unit being dropped. The form of linear constraints suggest that in the general case of Q applications, the feasible region is a convex polytope.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 14 Feasible Region - Example

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 15 A point p1 dominates a point p2 when for each application q, If the current system state is p2, one can apply the rate allocations calculated for p1 A Pareto point is not dominated by any other point in the feasible region Dominance & Pareto Points Pareto points represent optimal solutions: There is no point that is “better” than a pareto point

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 16 If the input rate of application q increases by, the input rate of a component of q will increase by In order for a stream processing system to be able to sustain such an increase, the following must hold for each node: Burst Handling Initial resource requirements Additional resource requirements due to single burst

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 17 Assuming simultaneous bursts, the capacity constraints for each node are: We assume each application has equal probability for a burst to appear. To minimize the amount of dropped data, we wish to maximize all : If the current input rates are represented by p, we need to configure the system for: Optimization Objective

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 18 Optimization Objective - Example To be able to accommodate the burst, the new point must be inside the feasible region

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 19 Our System: BARRE Our objective is to accommodate bursts in real- time through rate reconfiguration This problem is difficult to solve online Instead, our solution: –Offline: Pre-calculates a small number of component rate assignment plans during an offline phase (using the Feasible Region & Pareto Point machinery that we described) –Online: Monitors application incoming rates and resource availability during runtime When bursts occur, component rates are re-assigned, based on the pre-calculated plans

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 20 Feasible Region Determination We don’t compute the entire feasible region Instead, we find a set of Pareto points and use them to construct the appropriate component rate assignment on the event of a burst For those Pareto points we pre-calculate their optimal rate assignment These component rate assignments will be used to construct the appropriate component rate allocation on the event of a burst.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 21 Identifying Index Points - Example Step 1: Find the maximum possible rate for each application –Solve a max-flow problem, with the rates of all but one applications set to 0.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 22 Identifying Index Points - Example Step 2: Find the mid-point of the resulting plane and see how far it can go –Solve max-flow by also constraining direction

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 23 Identifying Index Points - Example Step 3: For each of the resulting sides, find each mid-point and repeat previous step –The upper left and lower right points are now dominated by the newly found index points

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 24 Identifying Index Points - Example Step 4: This is the resulting approximation to the feasible region –The mid-points of the sides cannot improve without breaking the capacity constraints

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems Online Phase Upon the onset of a burst, we use the pre- calculated Pareto Points to determine the appropriate component input rate allocation plan We may need to adjust the entire plan as a result of the burst 25 p’ = p +δ p 2 dominates p, however, when there is a burst we may have to change plans to p3 p1p1 p p3p3 p2p2 p’

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 26 Experimental Setup Implementation over our Synergy middleware Used FreePastry for service discovery and collection of statistics 10 unique services in the system, 5 services on a single node Each application included 4 to 6 services Each result is an average of 5 runs Comparison with: –A burst-unaware method –Static Reservation of 20% of node resources –Simple Dynamic Adaptation, from previous work –A combination of the above

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 27 Index Point Pre-Calculation Index Point (Pareto Point) Database calculation time, this is why index points need to be pre-calculated.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 28 BARRE Operation BARRE avoids missing data units during the burst and minimizes lost data units. Also, resulting plan is “safer”, so it prevents future drops

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 29 Missed Data Units BARRE results in fewer data units dropped. It can sustain up to about 80% (vs. about 20%) application rate increase without dropping any data unit.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 30 Data Units Delivered On Time BARRE achieves a high percentage of data unit delivery.

Yannis Drougas, Vana Kalogeraki Accommodating Bursts in Distributed Stream Processing Systems 31 Conclusions We proposed BARRE, a dynamic reservation scheme to address bursts in a dynamic stream processing system. Reservation is based on application needs and readjusted according to system dynamics BARRE utilizes multiple nodes for each processing component, splitting the input rate among them whenever needed

Accommodating Bursts in Distributed Stream Processing Systems Yannis Drougas, ESRI Vana Kalogeraki, AUEB