An Adaptive Load Balancing Middleware for Distributed Simulation
1 like684 views
The document presents the GAIA+ adaptive load balancing middleware for distributed simulations, aimed at improving simulation speed by dynamically reallocating simulated model entities. It addresses the challenges of communication and load balancing in parallel and distributed simulations, advocating for a migration-based approach to optimize resource utilization. Performance evaluations demonstrate that GAIA+ significantly outperforms static allocation schemes, showcasing its effectiveness in reducing wall clock time during simulations.
![Gabriele D’Angelo [email_address] http://www.cs.unibo.it/~gdangelo University of Bologna Department of Computer Science WOMP 2006, Sorrento (NA) An Adaptive Load Balancing Middleware for Distributed Simulation](https://image.slidesharecdn.com/an-adaptive-load-balancing-middleware-for-distributed-simulation-11779/85/An-Adaptive-Load-Balancing-Middleware-for-Distributed-Simulation-17-320.jpg)
An Adaptive Load Balancing Middleware for Distributed Simulation
- 1. Gabriele D’Angelo joint work with Luciano Bononi Michele Bracuto Lorenzo Donatiello University of Bologna Department of Computer Science WOMP 2006 – Sorrento (NA) An Adaptive Load Balancing Middleware for Distributed Simulation
- 2. Presentation outline Basic assumptions and goals Parallel and Distributed Simulation (PADS) main problems A migration-based approach to improve the simulation speed The proposed adaptive load-balancing mechanism Testbed: Ad Hoc network model Execution architecture definition Simulation results: dynamic reallocation and simulation speed Conclusions and future work
- 3. Basic assumptions and goal The simulation is a useful technique to support the design and the performance evaluation of complex systems The systems now considered of interest are composed by many highly dynamic entities, with unpredictable communication patterns The simulation of such kind of systems is possible only aggregating together many execution units: Parallel And Distributed Simulation (PADS) Our goal is to increase the simulation speed , that is reduce the Wall Clock Time (WCT) required to complete the simulation runs
- 4. Parallel and Distributed Simulation (PADS) PROS : Clustering together many Physical Execution Units (PEUs) it is possible to reduce the WCT required to complete a simulation Aggregating resources (i.e. RAM) it is possible to represent very complex systems CONS : A distributed simulation requires a large amount of communication and synchronization to obtain correct results The simulation speed depends on may factors, in example: heterogeneous CPUs and resources, background load
- 5. The GAIA migration-based framework A distributed simulation can be seen as a set of Logical Processes (LPs). Each LP is run by a possibly different PEU and takes care to manage the evolution of a set of Simulated Model Entities (SMEs) What is a good scheme to allocate the SMEs on the LPs/PEUs? We have demonstrated that the static allocation schemes are in most cases sub-optimal , and that a migration-based approach of the simulated model entities (GAIA) can Reduce the amount of communication and synchronization, clustering together the highly interacting simulated entities within the same LP/PEU Reduce the WCT, therefore increase the simulation speed-up
- 6. PADS execution architectures Dedicated clusters composed by homogeneous units are costly and often underutilized For the same reason, shared clusters composed by available Commercial-Off-the-Shelf (COTS) hardware are preferable to dedicated systems The simulation performance is highly influenced by the CPU load in background In a shared cluster it is impossible to predict a good static allocation scheme for the simulated entities: the background load is unpredictable and the CPU can usually highly heterogeneous An adaptive load balancing mechanism could improve the resources utilization and therefore the simulation speed
- 7. The GAIA+ adaptive load balancing middleware GAIA+ is an evolution of the migration-based mechanism GAIA It is composed by two cooperative parts, both based on the reallocation of simulated entities: The heuristic migration policy : to adapt and reduce the communication and synchronization needs The heuristic load-balancing policy : the overloaded (and therefore “slow”) PEUs can migrate some of the managed model entities to unloaded PEUs The simulation is totally distributed and therefore there is no point of centralization. Slow and Fast are attributes that have to be observed locally, with only a partial knowledge of the whole system and influenced by the networks delays
- 8. Performance evaluation: Ad-Hoc wireless network model The GAIA mechanism outperforms the static distributed simulation approach. This performance evaluation will compare the GAIA and the GAIA+ mechanisms Ad-Hoc wireless network model definition : 9000 Simulated Mobile Hosts (SMHs) over a flat topology Mobility model: Random Waypoint Model (RWP) uncorrelated SMHs’ mobility Traffic model: ping messages (CBR) by every SMH to all neighbors within the wireless communication range (250 m) Propagation model open space (neighbor-SMHs within detection range)
- 9. Ad-Hoc network model characterization Computation and communication issues: The computation required for each SMH per time-step is in the order of O(#SMH ^ 2) : to determine the neighbor set The communication required among SMHs is in the order of O(K*#SMH) per time-step, with K defined as a constant value based on SMHs density (assumed as constant)
- 10. Testbed execution architecture Distributed simulation, execution architecture: 3 heterogeneous PEU: 2 - Intel Dual Xeon Pentium IV 2800 MHz, with 3 GB RAM and 4 GB RAM , Debian GNU/Linux OS with kernel version 2.6.x 1 - Intel Quad Xeon Pentium IV 1500 MHz, with 1 GB RAM , Debian GNU/Linux OS with kernel version 2.6.x PEUs are interconnected by a Gigabit Ethernet LAN 3 LPs = 1 LP for each PEU Conservative time-stepped simulation: 2000 time-steps
- 11. Performance evaluation: three different scenarios The performance evaluation of the GAIA+ mechanism has involved three different scenarios: The CPU are considered as homogeneous: initially each PEU allocates the same number of SMHs The initial allocation is based on the nominal performance of the CPUs (as expressed in MHz) A synthetic background load (in form of a sinusoidal wave) is injected in part of the simulation cluster
- 12. Performance evaluation: Equal Distribution (scenario a)
- 13. Performance evaluation: MHz-based Distribution (scenario b)
- 14. Performance evaluation: Variable Background Load (scenario c)
- 15. Performance evaluation: Wall-Clock-Time (WCT) WCT (seconds) to complete a simulation run of 1000 timesteps -21.10 % 4128 5232 c -10.41 % 3568 3989 b -4.38 % 3442 3600 a diff (%) GAIA+ GAIA scenario
- 16. Conclusions and future work The communication and the load balancing are two main problems of PADS, they should be managed together An approach based on the migration of simulated entities can: reduce the communication overhead of a distributed simulation, induce a good level of load balancing and therefore increase the simulation speed In many cases the very costly High-Performance-Computing (HPC) clusters can be replaced by shared clusters of Commercial-Off-the-Shelf (COTS) computers The Computational Grid architecture could benefit of similar approaches, enhancing the use of the Grid for simulation tasks
- 17. Gabriele D’Angelo [email_address] http://www.cs.unibo.it/~gdangelo University of Bologna Department of Computer Science WOMP 2006, Sorrento (NA) An Adaptive Load Balancing Middleware for Distributed Simulation